diff --git "a/trainer_state.json" "b/trainer_state.json"
new file mode 100644--- /dev/null
+++ "b/trainer_state.json"
@@ -0,0 +1,52033 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.2631578947368421,
+  "eval_steps": 2000,
+  "global_step": 40000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 6.578947368421052e-05,
+      "grad_norm": 992.0,
+      "learning_rate": 1e-05,
+      "loss": 37.1063,
+      "loss/crossentropy": 15.088774585723877,
+      "loss/hidden": 19.0875,
+      "loss/incoh": 0.0,
+      "loss/logits": 17.867034912109375,
+      "loss/reg": 0.0,
+      "step": 10
+    },
+    {
+      "epoch": 0.00013157894736842105,
+      "grad_norm": 408.0,
+      "grad_norm_var": 138977.05,
+      "learning_rate": 2e-05,
+      "loss": 34.8921,
+      "loss/crossentropy": 14.647695541381836,
+      "loss/hidden": 18.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 15.686719226837159,
+      "loss/reg": 0.0,
+      "step": 20
+    },
+    {
+      "epoch": 0.00019736842105263157,
+      "grad_norm": 296.0,
+      "grad_norm_var": 17795.066666666666,
+      "learning_rate": 3e-05,
+      "loss": 32.2418,
+      "loss/crossentropy": 14.60380687713623,
+      "loss/hidden": 18.6375,
+      "loss/incoh": 0.0,
+      "loss/logits": 13.480152130126953,
+      "loss/reg": 0.0,
+      "step": 30
+    },
+    {
+      "epoch": 0.0002631578947368421,
+      "grad_norm": 26.125,
+      "grad_norm_var": 30349.038525390624,
+      "learning_rate": 4e-05,
+      "loss": 28.5558,
+      "loss/crossentropy": 15.434869766235352,
+      "loss/hidden": 18.525,
+      "loss/incoh": 0.0,
+      "loss/logits": 10.245227527618407,
+      "loss/reg": 0.0,
+      "step": 40
+    },
+    {
+      "epoch": 0.0003289473684210526,
+      "grad_norm": 51.25,
+      "grad_norm_var": 160.51717122395834,
+      "learning_rate": 5e-05,
+      "loss": 27.8535,
+      "loss/crossentropy": 13.159739780426026,
+      "loss/hidden": 18.4,
+      "loss/incoh": 0.0,
+      "loss/logits": 10.088242149353027,
+      "loss/reg": 0.0,
+      "step": 50
+    },
+    {
+      "epoch": 0.00039473684210526315,
+      "grad_norm": 22.125,
+      "grad_norm_var": 1.3315524858929522e+17,
+      "learning_rate": 6e-05,
+      "loss": 26.8127,
+      "loss/crossentropy": 10.868320941925049,
+      "loss/hidden": 18.1375,
+      "loss/incoh": 0.0,
+      "loss/logits": 8.643436527252197,
+      "loss/reg": 0.0,
+      "step": 60
+    },
+    {
+      "epoch": 0.0004605263157894737,
+      "grad_norm": 25.625,
+      "grad_norm_var": 1.3315524982997056e+17,
+      "learning_rate": 7e-05,
+      "loss": 25.6909,
+      "loss/crossentropy": 10.59145736694336,
+      "loss/hidden": 17.65,
+      "loss/incoh": 0.0,
+      "loss/logits": 8.497289371490478,
+      "loss/reg": 0.0,
+      "step": 70
+    },
+    {
+      "epoch": 0.0005263157894736842,
+      "grad_norm": 32.75,
+      "grad_norm_var": 206.156103515625,
+      "learning_rate": 8e-05,
+      "loss": 24.5919,
+      "loss/crossentropy": 10.117278575897217,
+      "loss/hidden": 17.025,
+      "loss/incoh": 0.0,
+      "loss/logits": 8.660355854034425,
+      "loss/reg": 0.0,
+      "step": 80
+    },
+    {
+      "epoch": 0.0005921052631578948,
+      "grad_norm": 128.0,
+      "grad_norm_var": 1129.743212890625,
+      "learning_rate": 9e-05,
+      "loss": 23.5218,
+      "loss/crossentropy": 9.528321361541748,
+      "loss/hidden": 16.3125,
+      "loss/incoh": 0.0,
+      "loss/logits": 6.695920991897583,
+      "loss/reg": 0.0,
+      "step": 90
+    },
+    {
+      "epoch": 0.0006578947368421052,
+      "grad_norm": 68.5,
+      "grad_norm_var": 439.56223958333334,
+      "learning_rate": 0.0001,
+      "loss": 23.2768,
+      "loss/crossentropy": 10.33118553161621,
+      "loss/hidden": 16.5125,
+      "loss/incoh": 0.0,
+      "loss/logits": 7.274227619171143,
+      "loss/reg": 0.0,
+      "step": 100
+    },
+    {
+      "epoch": 0.0007236842105263158,
+      "grad_norm": 73.5,
+      "grad_norm_var": 2741.439518229167,
+      "learning_rate": 0.0001,
+      "loss": 22.6559,
+      "loss/crossentropy": 9.676900672912598,
+      "loss/hidden": 16.21875,
+      "loss/incoh": 0.0,
+      "loss/logits": 7.22898006439209,
+      "loss/reg": 0.0,
+      "step": 110
+    },
+    {
+      "epoch": 0.0007894736842105263,
+      "grad_norm": 115.5,
+      "grad_norm_var": 1451.31875,
+      "learning_rate": 0.0001,
+      "loss": 22.7316,
+      "loss/crossentropy": 9.472201251983643,
+      "loss/hidden": 16.03125,
+      "loss/incoh": 0.0,
+      "loss/logits": 6.231936502456665,
+      "loss/reg": 0.0,
+      "step": 120
+    },
+    {
+      "epoch": 0.0008552631578947369,
+      "grad_norm": 32.25,
+      "grad_norm_var": 939.03515625,
+      "learning_rate": 0.0001,
+      "loss": 22.1075,
+      "loss/crossentropy": 9.909808540344239,
+      "loss/hidden": 16.075,
+      "loss/incoh": 0.0,
+      "loss/logits": 6.337067890167236,
+      "loss/reg": 0.0,
+      "step": 130
+    },
+    {
+      "epoch": 0.0009210526315789473,
+      "grad_norm": 37.75,
+      "grad_norm_var": 222.684375,
+      "learning_rate": 0.0001,
+      "loss": 22.103,
+      "loss/crossentropy": 9.516400051116943,
+      "loss/hidden": 15.8875,
+      "loss/incoh": 0.0,
+      "loss/logits": 5.677393054962158,
+      "loss/reg": 0.0,
+      "step": 140
+    },
+    {
+      "epoch": 0.000986842105263158,
+      "grad_norm": 46.25,
+      "grad_norm_var": 1221.1551432291667,
+      "learning_rate": 0.0001,
+      "loss": 21.5532,
+      "loss/crossentropy": 9.398476314544677,
+      "loss/hidden": 15.70625,
+      "loss/incoh": 0.0,
+      "loss/logits": 6.764282178878784,
+      "loss/reg": 0.0,
+      "step": 150
+    },
+    {
+      "epoch": 0.0010526315789473684,
+      "grad_norm": 42.0,
+      "grad_norm_var": 416.9809895833333,
+      "learning_rate": 0.0001,
+      "loss": 20.8381,
+      "loss/crossentropy": 9.098556232452392,
+      "loss/hidden": 15.225,
+      "loss/incoh": 0.0,
+      "loss/logits": 5.8288336277008055,
+      "loss/reg": 0.0,
+      "step": 160
+    },
+    {
+      "epoch": 0.0011184210526315789,
+      "grad_norm": 33.25,
+      "grad_norm_var": 334.8979166666667,
+      "learning_rate": 0.0001,
+      "loss": 18.9261,
+      "loss/crossentropy": 7.879999303817749,
+      "loss/hidden": 13.65,
+      "loss/incoh": 0.0,
+      "loss/logits": 5.353159737586975,
+      "loss/reg": 0.0,
+      "step": 170
+    },
+    {
+      "epoch": 0.0011842105263157896,
+      "grad_norm": 14.375,
+      "grad_norm_var": 250.13333333333333,
+      "learning_rate": 0.0001,
+      "loss": 16.5004,
+      "loss/crossentropy": 6.681199312210083,
+      "loss/hidden": 12.0125,
+      "loss/incoh": 0.0,
+      "loss/logits": 4.6115447044372555,
+      "loss/reg": 0.0,
+      "step": 180
+    },
+    {
+      "epoch": 0.00125,
+      "grad_norm": 12.5625,
+      "grad_norm_var": 119.9546875,
+      "learning_rate": 0.0001,
+      "loss": 14.1282,
+      "loss/crossentropy": 5.785468435287475,
+      "loss/hidden": 10.725,
+      "loss/incoh": 0.0,
+      "loss/logits": 3.489908790588379,
+      "loss/reg": 0.0,
+      "step": 190
+    },
+    {
+      "epoch": 0.0013157894736842105,
+      "grad_norm": 12.625,
+      "grad_norm_var": 32.467431640625,
+      "learning_rate": 0.0001,
+      "loss": 12.8216,
+      "loss/crossentropy": 4.923622274398804,
+      "loss/hidden": 9.675,
+      "loss/incoh": 0.0,
+      "loss/logits": 3.1496715903282166,
+      "loss/reg": 0.0,
+      "step": 200
+    },
+    {
+      "epoch": 0.001381578947368421,
+      "grad_norm": 26.375,
+      "grad_norm_var": 22.6759765625,
+      "learning_rate": 0.0001,
+      "loss": 11.5516,
+      "loss/crossentropy": 4.429650473594665,
+      "loss/hidden": 8.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 2.247162342071533,
+      "loss/reg": 0.0,
+      "step": 210
+    },
+    {
+      "epoch": 0.0014473684210526317,
+      "grad_norm": 35.5,
+      "grad_norm_var": 45.67233072916667,
+      "learning_rate": 0.0001,
+      "loss": 10.495,
+      "loss/crossentropy": 4.112493515014648,
+      "loss/hidden": 8.346875,
+      "loss/incoh": 0.0,
+      "loss/logits": 2.2232163310050965,
+      "loss/reg": 0.0,
+      "step": 220
+    },
+    {
+      "epoch": 0.0015131578947368421,
+      "grad_norm": 37.5,
+      "grad_norm_var": 58.01770833333333,
+      "learning_rate": 0.0001,
+      "loss": 9.9703,
+      "loss/crossentropy": 4.019938945770264,
+      "loss/hidden": 8.015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 2.0913679361343385,
+      "loss/reg": 0.0,
+      "step": 230
+    },
+    {
+      "epoch": 0.0015789473684210526,
+      "grad_norm": 36.25,
+      "grad_norm_var": 53.95390625,
+      "learning_rate": 0.0001,
+      "loss": 9.5394,
+      "loss/crossentropy": 3.9986461877822874,
+      "loss/hidden": 7.490625,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.801698899269104,
+      "loss/reg": 0.0,
+      "step": 240
+    },
+    {
+      "epoch": 0.001644736842105263,
+      "grad_norm": 34.0,
+      "grad_norm_var": 913.9712890625,
+      "learning_rate": 0.0001,
+      "loss": 9.1248,
+      "loss/crossentropy": 3.7174268484115602,
+      "loss/hidden": 7.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 2.0452203273773195,
+      "loss/reg": 0.0,
+      "step": 250
+    },
+    {
+      "epoch": 0.0017105263157894738,
+      "grad_norm": 30.625,
+      "grad_norm_var": 901.7176432291667,
+      "learning_rate": 0.0001,
+      "loss": 8.7525,
+      "loss/crossentropy": 3.6645556688308716,
+      "loss/hidden": 7.36875,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.4742069363594055,
+      "loss/reg": 0.0,
+      "step": 260
+    },
+    {
+      "epoch": 0.0017763157894736842,
+      "grad_norm": 40.0,
+      "grad_norm_var": 110.65826822916667,
+      "learning_rate": 0.0001,
+      "loss": 8.8113,
+      "loss/crossentropy": 3.116227722167969,
+      "loss/hidden": 7.31875,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.6629727721214294,
+      "loss/reg": 0.0,
+      "step": 270
+    },
+    {
+      "epoch": 0.0018421052631578947,
+      "grad_norm": 34.75,
+      "grad_norm_var": 122.43170572916667,
+      "learning_rate": 0.0001,
+      "loss": 8.5312,
+      "loss/crossentropy": 3.413420820236206,
+      "loss/hidden": 7.38125,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.3304585099220276,
+      "loss/reg": 0.0,
+      "step": 280
+    },
+    {
+      "epoch": 0.0019078947368421052,
+      "grad_norm": 30.25,
+      "grad_norm_var": 50.46640625,
+      "learning_rate": 0.0001,
+      "loss": 8.1172,
+      "loss/crossentropy": 3.313588786125183,
+      "loss/hidden": 6.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.1558095216751099,
+      "loss/reg": 0.0,
+      "step": 290
+    },
+    {
+      "epoch": 0.001973684210526316,
+      "grad_norm": 33.5,
+      "grad_norm_var": 66.95618489583333,
+      "learning_rate": 0.0001,
+      "loss": 8.4831,
+      "loss/crossentropy": 3.286371445655823,
+      "loss/hidden": 6.971875,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.4257299542427062,
+      "loss/reg": 0.0,
+      "step": 300
+    },
+    {
+      "epoch": 0.0020394736842105263,
+      "grad_norm": 37.0,
+      "grad_norm_var": 39.40930989583333,
+      "learning_rate": 0.0001,
+      "loss": 8.1428,
+      "loss/crossentropy": 3.1924397230148314,
+      "loss/hidden": 6.7375,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.129437392950058,
+      "loss/reg": 0.0,
+      "step": 310
+    },
+    {
+      "epoch": 0.002105263157894737,
+      "grad_norm": 35.75,
+      "grad_norm_var": 60.73125,
+      "learning_rate": 0.0001,
+      "loss": 8.1236,
+      "loss/crossentropy": 3.217240035533905,
+      "loss/hidden": 7.0375,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.216874635219574,
+      "loss/reg": 0.0,
+      "step": 320
+    },
+    {
+      "epoch": 0.0021710526315789473,
+      "grad_norm": 32.75,
+      "grad_norm_var": 21.7181640625,
+      "learning_rate": 0.0001,
+      "loss": 8.0115,
+      "loss/crossentropy": 3.2230591058731077,
+      "loss/hidden": 6.665625,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.165043205022812,
+      "loss/reg": 0.0,
+      "step": 330
+    },
+    {
+      "epoch": 0.0022368421052631577,
+      "grad_norm": 38.75,
+      "grad_norm_var": 247.53932291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.8717,
+      "loss/crossentropy": 3.491655874252319,
+      "loss/hidden": 6.628125,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.1553439140319823,
+      "loss/reg": 0.0,
+      "step": 340
+    },
+    {
+      "epoch": 0.002302631578947368,
+      "grad_norm": 29.0,
+      "grad_norm_var": 243.79576822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.8354,
+      "loss/crossentropy": 3.3709447622299193,
+      "loss/hidden": 6.646875,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.1052397668361664,
+      "loss/reg": 0.0,
+      "step": 350
+    },
+    {
+      "epoch": 0.002368421052631579,
+      "grad_norm": 28.75,
+      "grad_norm_var": 40.61640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5894,
+      "loss/crossentropy": 3.0430339336395265,
+      "loss/hidden": 6.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.152853137254715,
+      "loss/reg": 0.0,
+      "step": 360
+    },
+    {
+      "epoch": 0.0024342105263157896,
+      "grad_norm": 50.0,
+      "grad_norm_var": 58.61640625,
+      "learning_rate": 0.0001,
+      "loss": 7.7607,
+      "loss/crossentropy": 3.461497259140015,
+      "loss/hidden": 6.640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.2907899796962738,
+      "loss/reg": 0.0,
+      "step": 370
+    },
+    {
+      "epoch": 0.0025,
+      "grad_norm": 23.875,
+      "grad_norm_var": 50.36145833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5895,
+      "loss/crossentropy": 2.8183989763259887,
+      "loss/hidden": 6.36875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.9597792446613311,
+      "loss/reg": 0.0,
+      "step": 380
+    },
+    {
+      "epoch": 0.0025657894736842105,
+      "grad_norm": 22.75,
+      "grad_norm_var": 18.348958333333332,
+      "learning_rate": 0.0001,
+      "loss": 7.4024,
+      "loss/crossentropy": 3.0434406876564024,
+      "loss/hidden": 6.425,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.1219703614711762,
+      "loss/reg": 0.0,
+      "step": 390
+    },
+    {
+      "epoch": 0.002631578947368421,
+      "grad_norm": 18.75,
+      "grad_norm_var": 75.46451822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.4663,
+      "loss/crossentropy": 2.9813458204269407,
+      "loss/hidden": 6.334375,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.0157755613327026,
+      "loss/reg": 0.0,
+      "step": 400
+    },
+    {
+      "epoch": 0.0026973684210526315,
+      "grad_norm": 20.25,
+      "grad_norm_var": 14.660416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4425,
+      "loss/crossentropy": 3.030743360519409,
+      "loss/hidden": 6.23125,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.0403401851654053,
+      "loss/reg": 0.0,
+      "step": 410
+    },
+    {
+      "epoch": 0.002763157894736842,
+      "grad_norm": 19.75,
+      "grad_norm_var": 6.712239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.1935,
+      "loss/crossentropy": 3.044888973236084,
+      "loss/hidden": 6.1125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.920581477880478,
+      "loss/reg": 0.0,
+      "step": 420
+    },
+    {
+      "epoch": 0.002828947368421053,
+      "grad_norm": 18.625,
+      "grad_norm_var": 5.847330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.0053,
+      "loss/crossentropy": 3.2355963468551634,
+      "loss/hidden": 6.015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.9773828387260437,
+      "loss/reg": 0.0,
+      "step": 430
+    },
+    {
+      "epoch": 0.0028947368421052633,
+      "grad_norm": 21.875,
+      "grad_norm_var": 9.627067057291667,
+      "learning_rate": 0.0001,
+      "loss": 6.9973,
+      "loss/crossentropy": 3.3775979042053224,
+      "loss/hidden": 6.071875,
+      "loss/incoh": 0.0,
+      "loss/logits": 1.0445533573627472,
+      "loss/reg": 0.0,
+      "step": 440
+    },
+    {
+      "epoch": 0.0029605263157894738,
+      "grad_norm": 19.25,
+      "grad_norm_var": 10.170947265625,
+      "learning_rate": 0.0001,
+      "loss": 6.9686,
+      "loss/crossentropy": 3.0577521562576293,
+      "loss/hidden": 5.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.8946591019630432,
+      "loss/reg": 0.0,
+      "step": 450
+    },
+    {
+      "epoch": 0.0030263157894736843,
+      "grad_norm": 19.75,
+      "grad_norm_var": 5.329166666666667,
+      "learning_rate": 0.0001,
+      "loss": 6.8021,
+      "loss/crossentropy": 3.2570735692977903,
+      "loss/hidden": 5.678125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.860103166103363,
+      "loss/reg": 0.0,
+      "step": 460
+    },
+    {
+      "epoch": 0.0030921052631578947,
+      "grad_norm": 11.5,
+      "grad_norm_var": 5.882535807291666,
+      "learning_rate": 0.0001,
+      "loss": 6.6494,
+      "loss/crossentropy": 3.045071005821228,
+      "loss/hidden": 5.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.890488612651825,
+      "loss/reg": 0.0,
+      "step": 470
+    },
+    {
+      "epoch": 0.003157894736842105,
+      "grad_norm": 12.0,
+      "grad_norm_var": 3.778369140625,
+      "learning_rate": 0.0001,
+      "loss": 6.6399,
+      "loss/crossentropy": 2.955122375488281,
+      "loss/hidden": 5.559375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.7988932132720947,
+      "loss/reg": 0.0,
+      "step": 480
+    },
+    {
+      "epoch": 0.0032236842105263157,
+      "grad_norm": 14.625,
+      "grad_norm_var": 3.397509765625,
+      "learning_rate": 0.0001,
+      "loss": 6.6006,
+      "loss/crossentropy": 2.8290895342826845,
+      "loss/hidden": 5.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.8468542337417603,
+      "loss/reg": 0.0,
+      "step": 490
+    },
+    {
+      "epoch": 0.003289473684210526,
+      "grad_norm": 11.5,
+      "grad_norm_var": 4.100113932291666,
+      "learning_rate": 0.0001,
+      "loss": 6.4823,
+      "loss/crossentropy": 2.7418078184127808,
+      "loss/hidden": 5.665625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.7723756909370423,
+      "loss/reg": 0.0,
+      "step": 500
+    },
+    {
+      "epoch": 0.003355263157894737,
+      "grad_norm": 11.5625,
+      "grad_norm_var": 2.79375,
+      "learning_rate": 0.0001,
+      "loss": 6.4511,
+      "loss/crossentropy": 3.1031686782836916,
+      "loss/hidden": 5.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.9047020822763443,
+      "loss/reg": 0.0,
+      "step": 510
+    },
+    {
+      "epoch": 0.0034210526315789475,
+      "grad_norm": 10.625,
+      "grad_norm_var": 0.8618326822916667,
+      "learning_rate": 0.0001,
+      "loss": 6.3114,
+      "loss/crossentropy": 2.7031071186065674,
+      "loss/hidden": 5.3875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.7112044870853425,
+      "loss/reg": 0.0,
+      "step": 520
+    },
+    {
+      "epoch": 0.003486842105263158,
+      "grad_norm": 9.625,
+      "grad_norm_var": 94.71066080729166,
+      "learning_rate": 0.0001,
+      "loss": 6.2538,
+      "loss/crossentropy": 2.8632609844207764,
+      "loss/hidden": 5.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.8032145172357559,
+      "loss/reg": 0.0,
+      "step": 530
+    },
+    {
+      "epoch": 0.0035526315789473684,
+      "grad_norm": 12.0,
+      "grad_norm_var": 2.7383748372395833,
+      "learning_rate": 0.0001,
+      "loss": 6.1042,
+      "loss/crossentropy": 3.037100100517273,
+      "loss/hidden": 5.25625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.7572773277759552,
+      "loss/reg": 0.0,
+      "step": 540
+    },
+    {
+      "epoch": 0.003618421052631579,
+      "grad_norm": 10.9375,
+      "grad_norm_var": 2.5140584309895835,
+      "learning_rate": 0.0001,
+      "loss": 6.1353,
+      "loss/crossentropy": 2.979613184928894,
+      "loss/hidden": 5.196875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.7719000339508056,
+      "loss/reg": 0.0,
+      "step": 550
+    },
+    {
+      "epoch": 0.0036842105263157894,
+      "grad_norm": 8.0625,
+      "grad_norm_var": 1.6379557291666667,
+      "learning_rate": 0.0001,
+      "loss": 6.0637,
+      "loss/crossentropy": 2.687799036502838,
+      "loss/hidden": 5.51875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.9166353821754456,
+      "loss/reg": 0.0,
+      "step": 560
+    },
+    {
+      "epoch": 0.00375,
+      "grad_norm": 9.9375,
+      "grad_norm_var": 18.912353515625,
+      "learning_rate": 0.0001,
+      "loss": 6.0506,
+      "loss/crossentropy": 2.935390818119049,
+      "loss/hidden": 5.165625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.716228786110878,
+      "loss/reg": 0.0,
+      "step": 570
+    },
+    {
+      "epoch": 0.0038157894736842103,
+      "grad_norm": 7.25,
+      "grad_norm_var": 18.753641764322918,
+      "learning_rate": 0.0001,
+      "loss": 5.9637,
+      "loss/crossentropy": 2.7780107259750366,
+      "loss/hidden": 5.190625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.7067849993705749,
+      "loss/reg": 0.0,
+      "step": 580
+    },
+    {
+      "epoch": 0.0038815789473684212,
+      "grad_norm": 6.84375,
+      "grad_norm_var": 2.39361572265625,
+      "learning_rate": 0.0001,
+      "loss": 5.9361,
+      "loss/crossentropy": 3.0060938119888307,
+      "loss/hidden": 5.225,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.7271955192089081,
+      "loss/reg": 0.0,
+      "step": 590
+    },
+    {
+      "epoch": 0.003947368421052632,
+      "grad_norm": 7.6875,
+      "grad_norm_var": 0.75357666015625,
+      "learning_rate": 0.0001,
+      "loss": 5.7669,
+      "loss/crossentropy": 2.691058301925659,
+      "loss/hidden": 4.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.6389567136764527,
+      "loss/reg": 0.0,
+      "step": 600
+    },
+    {
+      "epoch": 0.004013157894736842,
+      "grad_norm": 6.78125,
+      "grad_norm_var": 2.53160400390625,
+      "learning_rate": 0.0001,
+      "loss": 5.7022,
+      "loss/crossentropy": 2.7504406690597536,
+      "loss/hidden": 5.121875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.8024603247642517,
+      "loss/reg": 0.0,
+      "step": 610
+    },
+    {
+      "epoch": 0.004078947368421053,
+      "grad_norm": 6.875,
+      "grad_norm_var": 2.7155558268229165,
+      "learning_rate": 0.0001,
+      "loss": 5.6145,
+      "loss/crossentropy": 2.9313748240470887,
+      "loss/hidden": 5.00625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.7257195949554444,
+      "loss/reg": 0.0,
+      "step": 620
+    },
+    {
+      "epoch": 0.0041447368421052636,
+      "grad_norm": 5.8125,
+      "grad_norm_var": 0.3809733072916667,
+      "learning_rate": 0.0001,
+      "loss": 5.6209,
+      "loss/crossentropy": 2.8686537384986877,
+      "loss/hidden": 5.05625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.7894174456596375,
+      "loss/reg": 0.0,
+      "step": 630
+    },
+    {
+      "epoch": 0.004210526315789474,
+      "grad_norm": 10.5,
+      "grad_norm_var": 1.7327962239583334,
+      "learning_rate": 0.0001,
+      "loss": 5.6566,
+      "loss/crossentropy": 2.910102880001068,
+      "loss/hidden": 4.95,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.7542287766933441,
+      "loss/reg": 0.0,
+      "step": 640
+    },
+    {
+      "epoch": 0.0042763157894736845,
+      "grad_norm": 6.0,
+      "grad_norm_var": 1.90406494140625,
+      "learning_rate": 0.0001,
+      "loss": 5.6548,
+      "loss/crossentropy": 3.0497835516929626,
+      "loss/hidden": 5.075,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.6385725855827331,
+      "loss/reg": 0.0,
+      "step": 650
+    },
+    {
+      "epoch": 0.0043421052631578945,
+      "grad_norm": 5.90625,
+      "grad_norm_var": 6.811181640625,
+      "learning_rate": 0.0001,
+      "loss": 5.4435,
+      "loss/crossentropy": 2.839945673942566,
+      "loss/hidden": 4.621875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.608047366142273,
+      "loss/reg": 0.0,
+      "step": 660
+    },
+    {
+      "epoch": 0.0044078947368421054,
+      "grad_norm": 8.5,
+      "grad_norm_var": 1.419775390625,
+      "learning_rate": 0.0001,
+      "loss": 5.4264,
+      "loss/crossentropy": 2.6664235949516297,
+      "loss/hidden": 4.71875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5786069691181183,
+      "loss/reg": 0.0,
+      "step": 670
+    },
+    {
+      "epoch": 0.0044736842105263155,
+      "grad_norm": 5.25,
+      "grad_norm_var": 1.222509765625,
+      "learning_rate": 0.0001,
+      "loss": 5.4175,
+      "loss/crossentropy": 2.7476831912994384,
+      "loss/hidden": 4.646875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.6524303257465363,
+      "loss/reg": 0.0,
+      "step": 680
+    },
+    {
+      "epoch": 0.004539473684210526,
+      "grad_norm": 5.15625,
+      "grad_norm_var": 0.8695271809895834,
+      "learning_rate": 0.0001,
+      "loss": 5.2974,
+      "loss/crossentropy": 2.718129062652588,
+      "loss/hidden": 4.69375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.679440614581108,
+      "loss/reg": 0.0,
+      "step": 690
+    },
+    {
+      "epoch": 0.004605263157894736,
+      "grad_norm": 5.875,
+      "grad_norm_var": 1.3825358072916667,
+      "learning_rate": 0.0001,
+      "loss": 5.3809,
+      "loss/crossentropy": 2.896076512336731,
+      "loss/hidden": 4.503125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.6036079049110412,
+      "loss/reg": 0.0,
+      "step": 700
+    },
+    {
+      "epoch": 0.004671052631578947,
+      "grad_norm": 5.875,
+      "grad_norm_var": 0.94888916015625,
+      "learning_rate": 0.0001,
+      "loss": 5.2593,
+      "loss/crossentropy": 2.765268421173096,
+      "loss/hidden": 4.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.7337387800216675,
+      "loss/reg": 0.0,
+      "step": 710
+    },
+    {
+      "epoch": 0.004736842105263158,
+      "grad_norm": 4.75,
+      "grad_norm_var": 1.0287760416666667,
+      "learning_rate": 0.0001,
+      "loss": 5.1503,
+      "loss/crossentropy": 2.6812595248222353,
+      "loss/hidden": 4.75625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.6710720509290695,
+      "loss/reg": 0.0,
+      "step": 720
+    },
+    {
+      "epoch": 0.004802631578947368,
+      "grad_norm": 6.0625,
+      "grad_norm_var": 86.4677734375,
+      "learning_rate": 0.0001,
+      "loss": 5.3144,
+      "loss/crossentropy": 2.7298573732376097,
+      "loss/hidden": 4.484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.6176893144845963,
+      "loss/reg": 0.0,
+      "step": 730
+    },
+    {
+      "epoch": 0.004868421052631579,
+      "grad_norm": 4.875,
+      "grad_norm_var": 85.56676025390625,
+      "learning_rate": 0.0001,
+      "loss": 5.131,
+      "loss/crossentropy": 2.823095703125,
+      "loss/hidden": 4.571875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.6797973781824111,
+      "loss/reg": 0.0,
+      "step": 740
+    },
+    {
+      "epoch": 0.004934210526315789,
+      "grad_norm": 4.96875,
+      "grad_norm_var": 4.8291015625,
+      "learning_rate": 0.0001,
+      "loss": 5.1654,
+      "loss/crossentropy": 2.901310992240906,
+      "loss/hidden": 4.61875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.737342044711113,
+      "loss/reg": 0.0,
+      "step": 750
+    },
+    {
+      "epoch": 0.005,
+      "grad_norm": 9.375,
+      "grad_norm_var": 3.796728515625,
+      "learning_rate": 0.0001,
+      "loss": 5.0448,
+      "loss/crossentropy": 2.5148804664611815,
+      "loss/hidden": 4.48125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5650010257959366,
+      "loss/reg": 0.0,
+      "step": 760
+    },
+    {
+      "epoch": 0.00506578947368421,
+      "grad_norm": 4.90625,
+      "grad_norm_var": 18.020947265625,
+      "learning_rate": 0.0001,
+      "loss": 5.0271,
+      "loss/crossentropy": 2.6732282817363737,
+      "loss/hidden": 4.48125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5763083070516586,
+      "loss/reg": 0.0,
+      "step": 770
+    },
+    {
+      "epoch": 0.005131578947368421,
+      "grad_norm": 4.71875,
+      "grad_norm_var": 17.617118326822915,
+      "learning_rate": 0.0001,
+      "loss": 5.0057,
+      "loss/crossentropy": 2.927682900428772,
+      "loss/hidden": 4.578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.680091741681099,
+      "loss/reg": 0.0,
+      "step": 780
+    },
+    {
+      "epoch": 0.005197368421052632,
+      "grad_norm": 4.53125,
+      "grad_norm_var": 1.5018880208333334,
+      "learning_rate": 0.0001,
+      "loss": 5.0918,
+      "loss/crossentropy": 2.7974375009536745,
+      "loss/hidden": 4.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.709694892168045,
+      "loss/reg": 0.0,
+      "step": 790
+    },
+    {
+      "epoch": 0.005263157894736842,
+      "grad_norm": 4.46875,
+      "grad_norm_var": 20.911812337239585,
+      "learning_rate": 0.0001,
+      "loss": 4.9367,
+      "loss/crossentropy": 2.875410461425781,
+      "loss/hidden": 4.578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.8118050575256348,
+      "loss/reg": 0.0,
+      "step": 800
+    },
+    {
+      "epoch": 0.005328947368421053,
+      "grad_norm": 5.375,
+      "grad_norm_var": 2.1762980143229167,
+      "learning_rate": 0.0001,
+      "loss": 4.8483,
+      "loss/crossentropy": 2.665932035446167,
+      "loss/hidden": 4.3140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5815477341413497,
+      "loss/reg": 0.0,
+      "step": 810
+    },
+    {
+      "epoch": 0.005394736842105263,
+      "grad_norm": 4.96875,
+      "grad_norm_var": 0.6493235270182292,
+      "learning_rate": 0.0001,
+      "loss": 4.9314,
+      "loss/crossentropy": 2.5820897936820986,
+      "loss/hidden": 4.15,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4888360023498535,
+      "loss/reg": 0.0,
+      "step": 820
+    },
+    {
+      "epoch": 0.005460526315789474,
+      "grad_norm": 3.890625,
+      "grad_norm_var": 0.2476470947265625,
+      "learning_rate": 0.0001,
+      "loss": 4.7859,
+      "loss/crossentropy": 2.647566497325897,
+      "loss/hidden": 4.3484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5885868102312088,
+      "loss/reg": 0.0,
+      "step": 830
+    },
+    {
+      "epoch": 0.005526315789473684,
+      "grad_norm": 4.9375,
+      "grad_norm_var": 0.29011128743489584,
+      "learning_rate": 0.0001,
+      "loss": 4.7829,
+      "loss/crossentropy": 2.583667039871216,
+      "loss/hidden": 4.2828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4921345829963684,
+      "loss/reg": 0.0,
+      "step": 840
+    },
+    {
+      "epoch": 0.005592105263157895,
+      "grad_norm": 4.6875,
+      "grad_norm_var": 0.397900390625,
+      "learning_rate": 0.0001,
+      "loss": 4.8821,
+      "loss/crossentropy": 2.798052990436554,
+      "loss/hidden": 4.2625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.7617209196090698,
+      "loss/reg": 0.0,
+      "step": 850
+    },
+    {
+      "epoch": 0.005657894736842106,
+      "grad_norm": 4.21875,
+      "grad_norm_var": 0.29501546223958336,
+      "learning_rate": 0.0001,
+      "loss": 4.9069,
+      "loss/crossentropy": 2.8075502276420594,
+      "loss/hidden": 4.1328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5111032873392105,
+      "loss/reg": 0.0,
+      "step": 860
+    },
+    {
+      "epoch": 0.005723684210526316,
+      "grad_norm": 4.21875,
+      "grad_norm_var": 0.24931233723958332,
+      "learning_rate": 0.0001,
+      "loss": 4.8049,
+      "loss/crossentropy": 2.400660240650177,
+      "loss/hidden": 4.3859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.6145752131938934,
+      "loss/reg": 0.0,
+      "step": 870
+    },
+    {
+      "epoch": 0.005789473684210527,
+      "grad_norm": 4.75,
+      "grad_norm_var": 1.2445271809895833,
+      "learning_rate": 0.0001,
+      "loss": 4.7333,
+      "loss/crossentropy": 2.631825530529022,
+      "loss/hidden": 4.3640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5737248510122299,
+      "loss/reg": 0.0,
+      "step": 880
+    },
+    {
+      "epoch": 0.005855263157894737,
+      "grad_norm": 4.90625,
+      "grad_norm_var": 1.9954742431640624,
+      "learning_rate": 0.0001,
+      "loss": 4.6882,
+      "loss/crossentropy": 2.9550926446914674,
+      "loss/hidden": 4.1046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5351347416639328,
+      "loss/reg": 0.0,
+      "step": 890
+    },
+    {
+      "epoch": 0.0059210526315789476,
+      "grad_norm": 4.15625,
+      "grad_norm_var": 0.8051991780598958,
+      "learning_rate": 0.0001,
+      "loss": 4.7267,
+      "loss/crossentropy": 2.322158467769623,
+      "loss/hidden": 4.2359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4901482403278351,
+      "loss/reg": 0.0,
+      "step": 900
+    },
+    {
+      "epoch": 0.005986842105263158,
+      "grad_norm": 4.59375,
+      "grad_norm_var": 0.5987589518229167,
+      "learning_rate": 0.0001,
+      "loss": 4.7552,
+      "loss/crossentropy": 2.163175904750824,
+      "loss/hidden": 4.475,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5075026541948319,
+      "loss/reg": 0.0,
+      "step": 910
+    },
+    {
+      "epoch": 0.0060526315789473685,
+      "grad_norm": 4.4375,
+      "grad_norm_var": 13.234305826822917,
+      "learning_rate": 0.0001,
+      "loss": 4.8996,
+      "loss/crossentropy": 2.5448178887367248,
+      "loss/hidden": 4.078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5253018319606781,
+      "loss/reg": 0.0,
+      "step": 920
+    },
+    {
+      "epoch": 0.0061184210526315785,
+      "grad_norm": 4.09375,
+      "grad_norm_var": 4.241829427083333,
+      "learning_rate": 0.0001,
+      "loss": 4.5704,
+      "loss/crossentropy": 2.551811099052429,
+      "loss/hidden": 4.0546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.48250589668750765,
+      "loss/reg": 0.0,
+      "step": 930
+    },
+    {
+      "epoch": 0.0061842105263157894,
+      "grad_norm": 4.71875,
+      "grad_norm_var": 0.13075764973958334,
+      "learning_rate": 0.0001,
+      "loss": 4.6287,
+      "loss/crossentropy": 2.8518677711486817,
+      "loss/hidden": 4.134375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5851545244455337,
+      "loss/reg": 0.0,
+      "step": 940
+    },
+    {
+      "epoch": 0.00625,
+      "grad_norm": 5.65625,
+      "grad_norm_var": 1.0635701497395833,
+      "learning_rate": 0.0001,
+      "loss": 4.6855,
+      "loss/crossentropy": 2.8685452222824095,
+      "loss/hidden": 3.9453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5299171417951584,
+      "loss/reg": 0.0,
+      "step": 950
+    },
+    {
+      "epoch": 0.00631578947368421,
+      "grad_norm": 4.4375,
+      "grad_norm_var": 0.2560506184895833,
+      "learning_rate": 0.0001,
+      "loss": 4.6193,
+      "loss/crossentropy": 2.73275808095932,
+      "loss/hidden": 4.121875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4992497324943542,
+      "loss/reg": 0.0,
+      "step": 960
+    },
+    {
+      "epoch": 0.006381578947368421,
+      "grad_norm": 4.0625,
+      "grad_norm_var": 0.3540598551432292,
+      "learning_rate": 0.0001,
+      "loss": 4.6436,
+      "loss/crossentropy": 2.6122008085250856,
+      "loss/hidden": 4.084375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5424144893884659,
+      "loss/reg": 0.0,
+      "step": 970
+    },
+    {
+      "epoch": 0.006447368421052631,
+      "grad_norm": 4.0,
+      "grad_norm_var": 1.2561260533134024e+17,
+      "learning_rate": 0.0001,
+      "loss": 4.7384,
+      "loss/crossentropy": 2.747082471847534,
+      "loss/hidden": 3.975,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4949415147304535,
+      "loss/reg": 0.0,
+      "step": 980
+    },
+    {
+      "epoch": 0.006513157894736842,
+      "grad_norm": 4.6875,
+      "grad_norm_var": 15.184098307291666,
+      "learning_rate": 0.0001,
+      "loss": 4.6154,
+      "loss/crossentropy": 2.6343679666519164,
+      "loss/hidden": 4.009375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.46223918795585633,
+      "loss/reg": 0.0,
+      "step": 990
+    },
+    {
+      "epoch": 0.006578947368421052,
+      "grad_norm": 4.84375,
+      "grad_norm_var": 3.58121337890625,
+      "learning_rate": 0.0001,
+      "loss": 4.6456,
+      "loss/crossentropy": 2.618730914592743,
+      "loss/hidden": 4.1390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.568024319410324,
+      "loss/reg": 0.0,
+      "step": 1000
+    },
+    {
+      "epoch": 0.006644736842105263,
+      "grad_norm": 10.1875,
+      "grad_norm_var": 2.6229563395182294,
+      "learning_rate": 0.0001,
+      "loss": 4.5241,
+      "loss/crossentropy": 2.7510436296463014,
+      "loss/hidden": 3.9421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4975563734769821,
+      "loss/reg": 0.0,
+      "step": 1010
+    },
+    {
+      "epoch": 0.006710526315789474,
+      "grad_norm": 4.59375,
+      "grad_norm_var": 9.821955362955729,
+      "learning_rate": 0.0001,
+      "loss": 4.5805,
+      "loss/crossentropy": 2.773310422897339,
+      "loss/hidden": 3.94375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5206439226865769,
+      "loss/reg": 0.0,
+      "step": 1020
+    },
+    {
+      "epoch": 0.006776315789473684,
+      "grad_norm": 4.34375,
+      "grad_norm_var": 8.482259114583334,
+      "learning_rate": 0.0001,
+      "loss": 4.4588,
+      "loss/crossentropy": 2.6046599745750427,
+      "loss/hidden": 3.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.44563083052635194,
+      "loss/reg": 0.0,
+      "step": 1030
+    },
+    {
+      "epoch": 0.006842105263157895,
+      "grad_norm": 4.0,
+      "grad_norm_var": 5.523110961914062,
+      "learning_rate": 0.0001,
+      "loss": 4.6377,
+      "loss/crossentropy": 2.6867428183555604,
+      "loss/hidden": 4.0921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4769616901874542,
+      "loss/reg": 0.0,
+      "step": 1040
+    },
+    {
+      "epoch": 0.006907894736842105,
+      "grad_norm": 3.890625,
+      "grad_norm_var": 9.615762329101562,
+      "learning_rate": 0.0001,
+      "loss": 4.5576,
+      "loss/crossentropy": 2.6374024391174316,
+      "loss/hidden": 3.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5235758543014526,
+      "loss/reg": 0.0,
+      "step": 1050
+    },
+    {
+      "epoch": 0.006973684210526316,
+      "grad_norm": 3.796875,
+      "grad_norm_var": 4.420035807291667,
+      "learning_rate": 0.0001,
+      "loss": 4.5061,
+      "loss/crossentropy": 2.409494662284851,
+      "loss/hidden": 4.0203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5403494209051132,
+      "loss/reg": 0.0,
+      "step": 1060
+    },
+    {
+      "epoch": 0.007039473684210526,
+      "grad_norm": 4.25,
+      "grad_norm_var": 12.098844401041667,
+      "learning_rate": 0.0001,
+      "loss": 4.7608,
+      "loss/crossentropy": 2.6496052145957947,
+      "loss/hidden": 3.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.48798912912607195,
+      "loss/reg": 0.0,
+      "step": 1070
+    },
+    {
+      "epoch": 0.007105263157894737,
+      "grad_norm": 3.875,
+      "grad_norm_var": 9.539867146809895,
+      "learning_rate": 0.0001,
+      "loss": 4.3815,
+      "loss/crossentropy": 2.757003378868103,
+      "loss/hidden": 4.11875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4590116262435913,
+      "loss/reg": 0.0,
+      "step": 1080
+    },
+    {
+      "epoch": 0.007171052631578947,
+      "grad_norm": 8.5625,
+      "grad_norm_var": 1.5923166910807292,
+      "learning_rate": 0.0001,
+      "loss": 4.44,
+      "loss/crossentropy": 2.550069880485535,
+      "loss/hidden": 4.071875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5490242928266525,
+      "loss/reg": 0.0,
+      "step": 1090
+    },
+    {
+      "epoch": 0.007236842105263158,
+      "grad_norm": 3.96875,
+      "grad_norm_var": 1.6922597249348958,
+      "learning_rate": 0.0001,
+      "loss": 4.4725,
+      "loss/crossentropy": 2.606226110458374,
+      "loss/hidden": 4.0875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5686961978673934,
+      "loss/reg": 0.0,
+      "step": 1100
+    },
+    {
+      "epoch": 0.007302631578947369,
+      "grad_norm": 4.90625,
+      "grad_norm_var": 0.9616282145182292,
+      "learning_rate": 0.0001,
+      "loss": 4.4416,
+      "loss/crossentropy": 2.742388653755188,
+      "loss/hidden": 3.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.491252401471138,
+      "loss/reg": 0.0,
+      "step": 1110
+    },
+    {
+      "epoch": 0.007368421052631579,
+      "grad_norm": 4.375,
+      "grad_norm_var": 0.8876302083333333,
+      "learning_rate": 0.0001,
+      "loss": 4.4119,
+      "loss/crossentropy": 2.858240842819214,
+      "loss/hidden": 3.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4695854902267456,
+      "loss/reg": 0.0,
+      "step": 1120
+    },
+    {
+      "epoch": 0.00743421052631579,
+      "grad_norm": 3.453125,
+      "grad_norm_var": 0.8083485921223958,
+      "learning_rate": 0.0001,
+      "loss": 4.3524,
+      "loss/crossentropy": 2.7758461236953735,
+      "loss/hidden": 3.90625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5061279594898224,
+      "loss/reg": 0.0,
+      "step": 1130
+    },
+    {
+      "epoch": 0.0075,
+      "grad_norm": 3.78125,
+      "grad_norm_var": 0.8257802327473959,
+      "learning_rate": 0.0001,
+      "loss": 4.296,
+      "loss/crossentropy": 2.849539041519165,
+      "loss/hidden": 3.7125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.45034482181072233,
+      "loss/reg": 0.0,
+      "step": 1140
+    },
+    {
+      "epoch": 0.007565789473684211,
+      "grad_norm": 4.0625,
+      "grad_norm_var": 0.22603759765625,
+      "learning_rate": 0.0001,
+      "loss": 4.259,
+      "loss/crossentropy": 2.6673625230789186,
+      "loss/hidden": 3.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.45979970395565034,
+      "loss/reg": 0.0,
+      "step": 1150
+    },
+    {
+      "epoch": 0.007631578947368421,
+      "grad_norm": 3.984375,
+      "grad_norm_var": 0.2490631103515625,
+      "learning_rate": 0.0001,
+      "loss": 4.2786,
+      "loss/crossentropy": 2.6413369178771973,
+      "loss/hidden": 3.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.45444311797618864,
+      "loss/reg": 0.0,
+      "step": 1160
+    },
+    {
+      "epoch": 0.007697368421052632,
+      "grad_norm": 6.1875,
+      "grad_norm_var": 0.7844960530598958,
+      "learning_rate": 0.0001,
+      "loss": 4.2937,
+      "loss/crossentropy": 2.54203085899353,
+      "loss/hidden": 3.6671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40531369894742963,
+      "loss/reg": 0.0,
+      "step": 1170
+    },
+    {
+      "epoch": 0.0077631578947368425,
+      "grad_norm": 3.65625,
+      "grad_norm_var": 0.444384765625,
+      "learning_rate": 0.0001,
+      "loss": 4.3125,
+      "loss/crossentropy": 2.772641682624817,
+      "loss/hidden": 3.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4890771210193634,
+      "loss/reg": 0.0,
+      "step": 1180
+    },
+    {
+      "epoch": 0.007828947368421053,
+      "grad_norm": 4.15625,
+      "grad_norm_var": 0.26546122233072916,
+      "learning_rate": 0.0001,
+      "loss": 4.3654,
+      "loss/crossentropy": 2.661901044845581,
+      "loss/hidden": 3.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.439369834959507,
+      "loss/reg": 0.0,
+      "step": 1190
+    },
+    {
+      "epoch": 0.007894736842105263,
+      "grad_norm": 3.75,
+      "grad_norm_var": 0.25321858723958335,
+      "learning_rate": 0.0001,
+      "loss": 4.2648,
+      "loss/crossentropy": 2.3847479939460756,
+      "loss/hidden": 4.028125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4770447015762329,
+      "loss/reg": 0.0,
+      "step": 1200
+    },
+    {
+      "epoch": 0.007960526315789473,
+      "grad_norm": 3.921875,
+      "grad_norm_var": 0.2809529622395833,
+      "learning_rate": 0.0001,
+      "loss": 4.3473,
+      "loss/crossentropy": 2.244320285320282,
+      "loss/hidden": 3.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40470985919237135,
+      "loss/reg": 0.0,
+      "step": 1210
+    },
+    {
+      "epoch": 0.008026315789473683,
+      "grad_norm": 4.28125,
+      "grad_norm_var": 0.5173004150390625,
+      "learning_rate": 0.0001,
+      "loss": 4.3494,
+      "loss/crossentropy": 2.5515334010124207,
+      "loss/hidden": 4.109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5172385692596435,
+      "loss/reg": 0.0,
+      "step": 1220
+    },
+    {
+      "epoch": 0.008092105263157895,
+      "grad_norm": 3.5,
+      "grad_norm_var": 0.2546539306640625,
+      "learning_rate": 0.0001,
+      "loss": 4.2293,
+      "loss/crossentropy": 2.5860470652580263,
+      "loss/hidden": 3.7171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4412991553544998,
+      "loss/reg": 0.0,
+      "step": 1230
+    },
+    {
+      "epoch": 0.008157894736842105,
+      "grad_norm": 3.1875,
+      "grad_norm_var": 1.7592185950961664e+17,
+      "learning_rate": 0.0001,
+      "loss": 4.4752,
+      "loss/crossentropy": 2.823768949508667,
+      "loss/hidden": 3.703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5228973954916001,
+      "loss/reg": 0.0,
+      "step": 1240
+    },
+    {
+      "epoch": 0.008223684210526315,
+      "grad_norm": 3.21875,
+      "grad_norm_var": 1.966552734375,
+      "learning_rate": 0.0001,
+      "loss": 4.2687,
+      "loss/crossentropy": 2.523781180381775,
+      "loss/hidden": 3.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4939111739397049,
+      "loss/reg": 0.0,
+      "step": 1250
+    },
+    {
+      "epoch": 0.008289473684210527,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 1.2621897379557292,
+      "learning_rate": 0.0001,
+      "loss": 4.1499,
+      "loss/crossentropy": 2.5173804640769957,
+      "loss/hidden": 3.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.44251940250396726,
+      "loss/reg": 0.0,
+      "step": 1260
+    },
+    {
+      "epoch": 0.008355263157894737,
+      "grad_norm": 4.15625,
+      "grad_norm_var": 3.665185546875,
+      "learning_rate": 0.0001,
+      "loss": 4.3378,
+      "loss/crossentropy": 2.551619827747345,
+      "loss/hidden": 3.9625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5328098922967911,
+      "loss/reg": 0.0,
+      "step": 1270
+    },
+    {
+      "epoch": 0.008421052631578947,
+      "grad_norm": 3.53125,
+      "grad_norm_var": 3.1853352864583333,
+      "learning_rate": 0.0001,
+      "loss": 4.2329,
+      "loss/crossentropy": 2.3984143674373626,
+      "loss/hidden": 3.9703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.43639505505561826,
+      "loss/reg": 0.0,
+      "step": 1280
+    },
+    {
+      "epoch": 0.008486842105263157,
+      "grad_norm": 4.21875,
+      "grad_norm_var": 0.21245015462239583,
+      "learning_rate": 0.0001,
+      "loss": 4.3604,
+      "loss/crossentropy": 2.8736027479171753,
+      "loss/hidden": 3.7875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.48700871765613557,
+      "loss/reg": 0.0,
+      "step": 1290
+    },
+    {
+      "epoch": 0.008552631578947369,
+      "grad_norm": 4.125,
+      "grad_norm_var": 0.24678446451822916,
+      "learning_rate": 0.0001,
+      "loss": 4.1666,
+      "loss/crossentropy": 2.714110541343689,
+      "loss/hidden": 3.75625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.47638387978076935,
+      "loss/reg": 0.0,
+      "step": 1300
+    },
+    {
+      "epoch": 0.008618421052631579,
+      "grad_norm": 5.03125,
+      "grad_norm_var": 0.36387430826822914,
+      "learning_rate": 0.0001,
+      "loss": 4.2879,
+      "loss/crossentropy": 2.6876192927360534,
+      "loss/hidden": 3.7,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.49314437210559847,
+      "loss/reg": 0.0,
+      "step": 1310
+    },
+    {
+      "epoch": 0.008684210526315789,
+      "grad_norm": 3.84375,
+      "grad_norm_var": 0.28189188639322915,
+      "learning_rate": 0.0001,
+      "loss": 4.1505,
+      "loss/crossentropy": 2.4753618359565737,
+      "loss/hidden": 3.684375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40235219299793246,
+      "loss/reg": 0.0,
+      "step": 1320
+    },
+    {
+      "epoch": 0.00875,
+      "grad_norm": 3.1875,
+      "grad_norm_var": 2.41715087890625,
+      "learning_rate": 0.0001,
+      "loss": 4.2095,
+      "loss/crossentropy": 2.515524423122406,
+      "loss/hidden": 3.996875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5540182292461395,
+      "loss/reg": 0.0,
+      "step": 1330
+    },
+    {
+      "epoch": 0.008815789473684211,
+      "grad_norm": 3.8125,
+      "grad_norm_var": 2.259993489583333,
+      "learning_rate": 0.0001,
+      "loss": 4.1889,
+      "loss/crossentropy": 2.346592426300049,
+      "loss/hidden": 3.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4610589429736137,
+      "loss/reg": 0.0,
+      "step": 1340
+    },
+    {
+      "epoch": 0.008881578947368421,
+      "grad_norm": 4.0625,
+      "grad_norm_var": 0.13492431640625,
+      "learning_rate": 0.0001,
+      "loss": 4.1587,
+      "loss/crossentropy": 2.600476896762848,
+      "loss/hidden": 3.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4497509777545929,
+      "loss/reg": 0.0,
+      "step": 1350
+    },
+    {
+      "epoch": 0.008947368421052631,
+      "grad_norm": 3.390625,
+      "grad_norm_var": 0.7162017822265625,
+      "learning_rate": 0.0001,
+      "loss": 4.0563,
+      "loss/crossentropy": 2.6253500103950502,
+      "loss/hidden": 3.5359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4007237285375595,
+      "loss/reg": 0.0,
+      "step": 1360
+    },
+    {
+      "epoch": 0.009013157894736843,
+      "grad_norm": 3.828125,
+      "grad_norm_var": 0.7642812093098958,
+      "learning_rate": 0.0001,
+      "loss": 4.1053,
+      "loss/crossentropy": 2.4306472063064577,
+      "loss/hidden": 3.596875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.42645111978054046,
+      "loss/reg": 0.0,
+      "step": 1370
+    },
+    {
+      "epoch": 0.009078947368421053,
+      "grad_norm": 5.65625,
+      "grad_norm_var": 0.3096995035807292,
+      "learning_rate": 0.0001,
+      "loss": 4.2093,
+      "loss/crossentropy": 2.4861367106437684,
+      "loss/hidden": 3.9359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.46326183080673217,
+      "loss/reg": 0.0,
+      "step": 1380
+    },
+    {
+      "epoch": 0.009144736842105263,
+      "grad_norm": 3.265625,
+      "grad_norm_var": 0.36946207682291665,
+      "learning_rate": 0.0001,
+      "loss": 4.0653,
+      "loss/crossentropy": 2.608224070072174,
+      "loss/hidden": 3.5140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3894644558429718,
+      "loss/reg": 0.0,
+      "step": 1390
+    },
+    {
+      "epoch": 0.009210526315789473,
+      "grad_norm": 3.21875,
+      "grad_norm_var": 0.30060221354166666,
+      "learning_rate": 0.0001,
+      "loss": 4.1547,
+      "loss/crossentropy": 2.337048816680908,
+      "loss/hidden": 3.6046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.39037723541259767,
+      "loss/reg": 0.0,
+      "step": 1400
+    },
+    {
+      "epoch": 0.009276315789473685,
+      "grad_norm": 3.625,
+      "grad_norm_var": 0.14612630208333333,
+      "learning_rate": 0.0001,
+      "loss": 4.076,
+      "loss/crossentropy": 2.726799726486206,
+      "loss/hidden": 3.5515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4537044405937195,
+      "loss/reg": 0.0,
+      "step": 1410
+    },
+    {
+      "epoch": 0.009342105263157895,
+      "grad_norm": 3.875,
+      "grad_norm_var": 1.9072662353515626,
+      "learning_rate": 0.0001,
+      "loss": 4.2387,
+      "loss/crossentropy": 2.5365243434906004,
+      "loss/hidden": 3.503125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37743023335933684,
+      "loss/reg": 0.0,
+      "step": 1420
+    },
+    {
+      "epoch": 0.009407894736842105,
+      "grad_norm": 4.1875,
+      "grad_norm_var": 0.15405171712239582,
+      "learning_rate": 0.0001,
+      "loss": 4.0913,
+      "loss/crossentropy": 2.6032602190971375,
+      "loss/hidden": 3.6015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4031028777360916,
+      "loss/reg": 0.0,
+      "step": 1430
+    },
+    {
+      "epoch": 0.009473684210526316,
+      "grad_norm": 3.453125,
+      "grad_norm_var": 0.35299072265625,
+      "learning_rate": 0.0001,
+      "loss": 4.0676,
+      "loss/crossentropy": 2.279827582836151,
+      "loss/hidden": 3.671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38540517538785934,
+      "loss/reg": 0.0,
+      "step": 1440
+    },
+    {
+      "epoch": 0.009539473684210526,
+      "grad_norm": 3.703125,
+      "grad_norm_var": 0.44612223307291665,
+      "learning_rate": 0.0001,
+      "loss": 4.0791,
+      "loss/crossentropy": 2.4795989274978636,
+      "loss/hidden": 3.7,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40130155086517333,
+      "loss/reg": 0.0,
+      "step": 1450
+    },
+    {
+      "epoch": 0.009605263157894737,
+      "grad_norm": 3.515625,
+      "grad_norm_var": 0.43585611979166666,
+      "learning_rate": 0.0001,
+      "loss": 4.0499,
+      "loss/crossentropy": 2.337530755996704,
+      "loss/hidden": 3.4703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3699365258216858,
+      "loss/reg": 0.0,
+      "step": 1460
+    },
+    {
+      "epoch": 0.009671052631578947,
+      "grad_norm": 4.375,
+      "grad_norm_var": 2.1455393473307294,
+      "learning_rate": 0.0001,
+      "loss": 4.0616,
+      "loss/crossentropy": 2.246569663286209,
+      "loss/hidden": 3.54375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37747917622327803,
+      "loss/reg": 0.0,
+      "step": 1470
+    },
+    {
+      "epoch": 0.009736842105263158,
+      "grad_norm": 4.34375,
+      "grad_norm_var": 1.3469017374654464e+17,
+      "learning_rate": 0.0001,
+      "loss": 4.1919,
+      "loss/crossentropy": 2.465463387966156,
+      "loss/hidden": 3.5796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.409694692492485,
+      "loss/reg": 0.0,
+      "step": 1480
+    },
+    {
+      "epoch": 0.009802631578947368,
+      "grad_norm": 3.8125,
+      "grad_norm_var": 2.647081560180774e+17,
+      "learning_rate": 0.0001,
+      "loss": 4.2391,
+      "loss/crossentropy": 2.614275646209717,
+      "loss/hidden": 3.86875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4961456567049026,
+      "loss/reg": 0.0,
+      "step": 1490
+    },
+    {
+      "epoch": 0.009868421052631578,
+      "grad_norm": 3.59375,
+      "grad_norm_var": 3.232743326822917,
+      "learning_rate": 0.0001,
+      "loss": 4.2432,
+      "loss/crossentropy": 2.9305615186691285,
+      "loss/hidden": 3.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.6574043720960617,
+      "loss/reg": 0.0,
+      "step": 1500
+    },
+    {
+      "epoch": 0.00993421052631579,
+      "grad_norm": 3.34375,
+      "grad_norm_var": 3.299430338541667,
+      "learning_rate": 0.0001,
+      "loss": 4.066,
+      "loss/crossentropy": 2.3778577923774717,
+      "loss/hidden": 3.6515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3975479930639267,
+      "loss/reg": 0.0,
+      "step": 1510
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 3.515625,
+      "grad_norm_var": 18.948729451497396,
+      "learning_rate": 0.0001,
+      "loss": 4.1724,
+      "loss/crossentropy": 2.527243709564209,
+      "loss/hidden": 3.596875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.41472980976104734,
+      "loss/reg": 0.0,
+      "step": 1520
+    },
+    {
+      "epoch": 0.01006578947368421,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 18.10924072265625,
+      "learning_rate": 0.0001,
+      "loss": 4.0302,
+      "loss/crossentropy": 2.7474317073822023,
+      "loss/hidden": 3.690625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.47478381991386415,
+      "loss/reg": 0.0,
+      "step": 1530
+    },
+    {
+      "epoch": 0.01013157894736842,
+      "grad_norm": 4.4375,
+      "grad_norm_var": 0.45779520670572915,
+      "learning_rate": 0.0001,
+      "loss": 4.0726,
+      "loss/crossentropy": 2.4822750091552734,
+      "loss/hidden": 3.48125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.39128718376159666,
+      "loss/reg": 0.0,
+      "step": 1540
+    },
+    {
+      "epoch": 0.010197368421052632,
+      "grad_norm": 3.5625,
+      "grad_norm_var": 0.21116129557291666,
+      "learning_rate": 0.0001,
+      "loss": 4.1353,
+      "loss/crossentropy": 2.480714201927185,
+      "loss/hidden": 3.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4798941880464554,
+      "loss/reg": 0.0,
+      "step": 1550
+    },
+    {
+      "epoch": 0.010263157894736842,
+      "grad_norm": 3.78125,
+      "grad_norm_var": 0.21678059895833332,
+      "learning_rate": 0.0001,
+      "loss": 3.993,
+      "loss/crossentropy": 2.536018407344818,
+      "loss/hidden": 3.53125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3731235474348068,
+      "loss/reg": 0.0,
+      "step": 1560
+    },
+    {
+      "epoch": 0.010328947368421052,
+      "grad_norm": 3.953125,
+      "grad_norm_var": 0.17219645182291668,
+      "learning_rate": 0.0001,
+      "loss": 4.0545,
+      "loss/crossentropy": 2.4370301008224486,
+      "loss/hidden": 3.5640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4435511589050293,
+      "loss/reg": 0.0,
+      "step": 1570
+    },
+    {
+      "epoch": 0.010394736842105264,
+      "grad_norm": 3.140625,
+      "grad_norm_var": 0.5761027018229167,
+      "learning_rate": 0.0001,
+      "loss": 4.0907,
+      "loss/crossentropy": 2.576425087451935,
+      "loss/hidden": 3.6484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.43103125393390657,
+      "loss/reg": 0.0,
+      "step": 1580
+    },
+    {
+      "epoch": 0.010460526315789474,
+      "grad_norm": 3.34375,
+      "grad_norm_var": 0.15403544108072917,
+      "learning_rate": 0.0001,
+      "loss": 3.8713,
+      "loss/crossentropy": 2.2120620369911195,
+      "loss/hidden": 3.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.43894679844379425,
+      "loss/reg": 0.0,
+      "step": 1590
+    },
+    {
+      "epoch": 0.010526315789473684,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.09127197265625,
+      "learning_rate": 0.0001,
+      "loss": 3.9719,
+      "loss/crossentropy": 2.5636333346366884,
+      "loss/hidden": 3.575,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4157550185918808,
+      "loss/reg": 0.0,
+      "step": 1600
+    },
+    {
+      "epoch": 0.010592105263157894,
+      "grad_norm": 3.5,
+      "grad_norm_var": 2.413444010416667,
+      "learning_rate": 0.0001,
+      "loss": 4.153,
+      "loss/crossentropy": 2.6213369131088258,
+      "loss/hidden": 3.6171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.45062357783317564,
+      "loss/reg": 0.0,
+      "step": 1610
+    },
+    {
+      "epoch": 0.010657894736842106,
+      "grad_norm": 9.5,
+      "grad_norm_var": 4.2955881754557295,
+      "learning_rate": 0.0001,
+      "loss": 4.0295,
+      "loss/crossentropy": 2.626167094707489,
+      "loss/hidden": 3.5765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.44043630063533784,
+      "loss/reg": 0.0,
+      "step": 1620
+    },
+    {
+      "epoch": 0.010723684210526316,
+      "grad_norm": 3.6875,
+      "grad_norm_var": 2.3217437744140623,
+      "learning_rate": 0.0001,
+      "loss": 3.937,
+      "loss/crossentropy": 2.6269264578819276,
+      "loss/hidden": 3.45625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.48427494168281554,
+      "loss/reg": 0.0,
+      "step": 1630
+    },
+    {
+      "epoch": 0.010789473684210526,
+      "grad_norm": 3.078125,
+      "grad_norm_var": 0.12195638020833334,
+      "learning_rate": 0.0001,
+      "loss": 3.9365,
+      "loss/crossentropy": 2.6117714166641237,
+      "loss/hidden": 3.4359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4069202274084091,
+      "loss/reg": 0.0,
+      "step": 1640
+    },
+    {
+      "epoch": 0.010855263157894738,
+      "grad_norm": 3.3125,
+      "grad_norm_var": 0.25017903645833334,
+      "learning_rate": 0.0001,
+      "loss": 3.9318,
+      "loss/crossentropy": 2.6508745312690736,
+      "loss/hidden": 3.53125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4166009187698364,
+      "loss/reg": 0.0,
+      "step": 1650
+    },
+    {
+      "epoch": 0.010921052631578948,
+      "grad_norm": 4.21875,
+      "grad_norm_var": 0.1591217041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.8964,
+      "loss/crossentropy": 2.4683817744255068,
+      "loss/hidden": 3.4421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37454236298799515,
+      "loss/reg": 0.0,
+      "step": 1660
+    },
+    {
+      "epoch": 0.010986842105263158,
+      "grad_norm": 3.28125,
+      "grad_norm_var": 0.12337239583333333,
+      "learning_rate": 0.0001,
+      "loss": 4.0102,
+      "loss/crossentropy": 2.4436564683914184,
+      "loss/hidden": 3.3390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36218023002147676,
+      "loss/reg": 0.0,
+      "step": 1670
+    },
+    {
+      "epoch": 0.011052631578947368,
+      "grad_norm": 3.15625,
+      "grad_norm_var": 0.08958231608072917,
+      "learning_rate": 0.0001,
+      "loss": 3.9784,
+      "loss/crossentropy": 2.559529435634613,
+      "loss/hidden": 3.703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.46595812439918516,
+      "loss/reg": 0.0,
+      "step": 1680
+    },
+    {
+      "epoch": 0.01111842105263158,
+      "grad_norm": 2.953125,
+      "grad_norm_var": 0.11448160807291667,
+      "learning_rate": 0.0001,
+      "loss": 3.9419,
+      "loss/crossentropy": 2.4433623433113096,
+      "loss/hidden": 3.453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.41454428136348725,
+      "loss/reg": 0.0,
+      "step": 1690
+    },
+    {
+      "epoch": 0.01118421052631579,
+      "grad_norm": 4.0625,
+      "grad_norm_var": 0.11336161295572916,
+      "learning_rate": 0.0001,
+      "loss": 3.8878,
+      "loss/crossentropy": 2.561389684677124,
+      "loss/hidden": 3.4640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3884895950555801,
+      "loss/reg": 0.0,
+      "step": 1700
+    },
+    {
+      "epoch": 0.01125,
+      "grad_norm": 3.296875,
+      "grad_norm_var": 2.4417928059895835,
+      "learning_rate": 0.0001,
+      "loss": 3.9996,
+      "loss/crossentropy": 2.709191393852234,
+      "loss/hidden": 3.5515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3691831022500992,
+      "loss/reg": 0.0,
+      "step": 1710
+    },
+    {
+      "epoch": 0.011315789473684211,
+      "grad_norm": 3.3125,
+      "grad_norm_var": 0.5698232014973958,
+      "learning_rate": 0.0001,
+      "loss": 4.0535,
+      "loss/crossentropy": 2.4276717066764832,
+      "loss/hidden": 3.7375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4981458902359009,
+      "loss/reg": 0.0,
+      "step": 1720
+    },
+    {
+      "epoch": 0.011381578947368421,
+      "grad_norm": 3.71875,
+      "grad_norm_var": 0.4803059895833333,
+      "learning_rate": 0.0001,
+      "loss": 4.0546,
+      "loss/crossentropy": 2.672085237503052,
+      "loss/hidden": 3.465625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.41347330510616304,
+      "loss/reg": 0.0,
+      "step": 1730
+    },
+    {
+      "epoch": 0.011447368421052631,
+      "grad_norm": 6.59375,
+      "grad_norm_var": 0.7734700520833333,
+      "learning_rate": 0.0001,
+      "loss": 4.0697,
+      "loss/crossentropy": 2.625990152359009,
+      "loss/hidden": 3.60625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.45684492886066436,
+      "loss/reg": 0.0,
+      "step": 1740
+    },
+    {
+      "epoch": 0.011513157894736841,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 2.146930948893229,
+      "learning_rate": 0.0001,
+      "loss": 4.1129,
+      "loss/crossentropy": 2.682978630065918,
+      "loss/hidden": 3.5703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4306509166955948,
+      "loss/reg": 0.0,
+      "step": 1750
+    },
+    {
+      "epoch": 0.011578947368421053,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.1972076416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.8797,
+      "loss/crossentropy": 2.7138221740722654,
+      "loss/hidden": 3.60625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.46029032766819,
+      "loss/reg": 0.0,
+      "step": 1760
+    },
+    {
+      "epoch": 0.011644736842105263,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 15.614777628580729,
+      "learning_rate": 0.0001,
+      "loss": 4.0409,
+      "loss/crossentropy": 2.1841426372528074,
+      "loss/hidden": 3.3828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35761781185865404,
+      "loss/reg": 0.0,
+      "step": 1770
+    },
+    {
+      "epoch": 0.011710526315789473,
+      "grad_norm": 3.953125,
+      "grad_norm_var": 0.32083231608072915,
+      "learning_rate": 0.0001,
+      "loss": 4.0523,
+      "loss/crossentropy": 2.404168051481247,
+      "loss/hidden": 3.325,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34855909645557404,
+      "loss/reg": 0.0,
+      "step": 1780
+    },
+    {
+      "epoch": 0.011776315789473683,
+      "grad_norm": 3.5,
+      "grad_norm_var": 3.124772135416667,
+      "learning_rate": 0.0001,
+      "loss": 3.9895,
+      "loss/crossentropy": 2.370392310619354,
+      "loss/hidden": 3.6515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4228974744677544,
+      "loss/reg": 0.0,
+      "step": 1790
+    },
+    {
+      "epoch": 0.011842105263157895,
+      "grad_norm": 6.125,
+      "grad_norm_var": 3.445572916666667,
+      "learning_rate": 0.0001,
+      "loss": 4.0349,
+      "loss/crossentropy": 2.6190654158592226,
+      "loss/hidden": 3.4578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.406630203127861,
+      "loss/reg": 0.0,
+      "step": 1800
+    },
+    {
+      "epoch": 0.011907894736842105,
+      "grad_norm": 3.28125,
+      "grad_norm_var": 0.575048828125,
+      "learning_rate": 0.0001,
+      "loss": 3.8828,
+      "loss/crossentropy": 2.505708968639374,
+      "loss/hidden": 3.4203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3773295432329178,
+      "loss/reg": 0.0,
+      "step": 1810
+    },
+    {
+      "epoch": 0.011973684210526315,
+      "grad_norm": 3.734375,
+      "grad_norm_var": 14.053641764322917,
+      "learning_rate": 0.0001,
+      "loss": 3.9952,
+      "loss/crossentropy": 2.8797521352767945,
+      "loss/hidden": 3.6734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.7419060736894607,
+      "loss/reg": 0.0,
+      "step": 1820
+    },
+    {
+      "epoch": 0.012039473684210527,
+      "grad_norm": 3.96875,
+      "grad_norm_var": 2.006696573893229,
+      "learning_rate": 0.0001,
+      "loss": 3.87,
+      "loss/crossentropy": 2.792651188373566,
+      "loss/hidden": 3.6546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.7776896879076958,
+      "loss/reg": 0.0,
+      "step": 1830
+    },
+    {
+      "epoch": 0.012105263157894737,
+      "grad_norm": 4.09375,
+      "grad_norm_var": 3.2296132405598956,
+      "learning_rate": 0.0001,
+      "loss": 3.8688,
+      "loss/crossentropy": 2.4713597655296327,
+      "loss/hidden": 3.4484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3962660849094391,
+      "loss/reg": 0.0,
+      "step": 1840
+    },
+    {
+      "epoch": 0.012171052631578947,
+      "grad_norm": 3.296875,
+      "grad_norm_var": 0.20377197265625,
+      "learning_rate": 0.0001,
+      "loss": 3.9721,
+      "loss/crossentropy": 2.225203812122345,
+      "loss/hidden": 3.4734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3860040009021759,
+      "loss/reg": 0.0,
+      "step": 1850
+    },
+    {
+      "epoch": 0.012236842105263157,
+      "grad_norm": 4.375,
+      "grad_norm_var": 14.807112630208334,
+      "learning_rate": 0.0001,
+      "loss": 4.0342,
+      "loss/crossentropy": 2.405521821975708,
+      "loss/hidden": 3.428125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40949456989765165,
+      "loss/reg": 0.0,
+      "step": 1860
+    },
+    {
+      "epoch": 0.012302631578947369,
+      "grad_norm": 4.53125,
+      "grad_norm_var": 6.056012980143229,
+      "learning_rate": 0.0001,
+      "loss": 3.9802,
+      "loss/crossentropy": 2.3927958846092223,
+      "loss/hidden": 3.5265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3984386846423149,
+      "loss/reg": 0.0,
+      "step": 1870
+    },
+    {
+      "epoch": 0.012368421052631579,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.6073893229166667,
+      "learning_rate": 0.0001,
+      "loss": 3.9074,
+      "loss/crossentropy": 2.6031975388526916,
+      "loss/hidden": 3.38125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4128950208425522,
+      "loss/reg": 0.0,
+      "step": 1880
+    },
+    {
+      "epoch": 0.012434210526315789,
+      "grad_norm": 3.65625,
+      "grad_norm_var": 0.3526845296223958,
+      "learning_rate": 0.0001,
+      "loss": 3.9143,
+      "loss/crossentropy": 2.7405603647232057,
+      "loss/hidden": 3.25625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37195596396923064,
+      "loss/reg": 0.0,
+      "step": 1890
+    },
+    {
+      "epoch": 0.0125,
+      "grad_norm": 3.3125,
+      "grad_norm_var": 130.27611389160157,
+      "learning_rate": 0.0001,
+      "loss": 3.986,
+      "loss/crossentropy": 2.6554584980010985,
+      "loss/hidden": 3.5,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3823524177074432,
+      "loss/reg": 0.0,
+      "step": 1900
+    },
+    {
+      "epoch": 0.01256578947368421,
+      "grad_norm": 3.6875,
+      "grad_norm_var": 130.43673400878907,
+      "learning_rate": 0.0001,
+      "loss": 3.8682,
+      "loss/crossentropy": 2.6816349744796755,
+      "loss/hidden": 3.421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.43074882328510283,
+      "loss/reg": 0.0,
+      "step": 1910
+    },
+    {
+      "epoch": 0.01263157894736842,
+      "grad_norm": 3.0,
+      "grad_norm_var": 10.65947265625,
+      "learning_rate": 0.0001,
+      "loss": 4.0152,
+      "loss/crossentropy": 2.1248778343200683,
+      "loss/hidden": 3.6953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4099419146776199,
+      "loss/reg": 0.0,
+      "step": 1920
+    },
+    {
+      "epoch": 0.01269736842105263,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.41845296223958334,
+      "learning_rate": 0.0001,
+      "loss": 3.9099,
+      "loss/crossentropy": 2.485488569736481,
+      "loss/hidden": 3.3921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3766200736165047,
+      "loss/reg": 0.0,
+      "step": 1930
+    },
+    {
+      "epoch": 0.012763157894736843,
+      "grad_norm": 4.375,
+      "grad_norm_var": 2.3640777587890627,
+      "learning_rate": 0.0001,
+      "loss": 4.0886,
+      "loss/crossentropy": 2.896865522861481,
+      "loss/hidden": 4.4,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.6787679702043533,
+      "loss/reg": 0.0,
+      "step": 1940
+    },
+    {
+      "epoch": 0.012828947368421053,
+      "grad_norm": 3.34375,
+      "grad_norm_var": 34.95194905598958,
+      "learning_rate": 0.0001,
+      "loss": 4.0285,
+      "loss/crossentropy": 2.348608684539795,
+      "loss/hidden": 3.6171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4279856622219086,
+      "loss/reg": 0.0,
+      "step": 1950
+    },
+    {
+      "epoch": 0.012894736842105263,
+      "grad_norm": 3.65625,
+      "grad_norm_var": 1.5377278645833334,
+      "learning_rate": 0.0001,
+      "loss": 3.9097,
+      "loss/crossentropy": 2.3533430814743044,
+      "loss/hidden": 3.48125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37981766164302827,
+      "loss/reg": 0.0,
+      "step": 1960
+    },
+    {
+      "epoch": 0.012960526315789474,
+      "grad_norm": 4.15625,
+      "grad_norm_var": 0.23593648274739584,
+      "learning_rate": 0.0001,
+      "loss": 3.9802,
+      "loss/crossentropy": 2.54624525308609,
+      "loss/hidden": 3.790625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4486588716506958,
+      "loss/reg": 0.0,
+      "step": 1970
+    },
+    {
+      "epoch": 0.013026315789473684,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.24025777180989583,
+      "learning_rate": 0.0001,
+      "loss": 3.8942,
+      "loss/crossentropy": 2.5797088146209717,
+      "loss/hidden": 3.7875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4970128297805786,
+      "loss/reg": 0.0,
+      "step": 1980
+    },
+    {
+      "epoch": 0.013092105263157894,
+      "grad_norm": 3.9375,
+      "grad_norm_var": 0.47609049479166665,
+      "learning_rate": 0.0001,
+      "loss": 3.934,
+      "loss/crossentropy": 2.555588161945343,
+      "loss/hidden": 3.38125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38218972086906433,
+      "loss/reg": 0.0,
+      "step": 1990
+    },
+    {
+      "epoch": 0.013157894736842105,
+      "grad_norm": 4.03125,
+      "grad_norm_var": 1.0361073811848958,
+      "learning_rate": 0.0001,
+      "loss": 3.8759,
+      "loss/crossentropy": 2.164474868774414,
+      "loss/hidden": 3.7125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3994155451655388,
+      "loss/reg": 0.0,
+      "step": 2000
+    },
+    {
+      "epoch": 0.013223684210526316,
+      "grad_norm": 3.453125,
+      "grad_norm_var": 0.8633778889973959,
+      "learning_rate": 0.0001,
+      "loss": 3.8317,
+      "loss/crossentropy": 2.558875060081482,
+      "loss/hidden": 3.471875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3639406472444534,
+      "loss/reg": 0.0,
+      "step": 2010
+    },
+    {
+      "epoch": 0.013289473684210526,
+      "grad_norm": 3.75,
+      "grad_norm_var": 1.5050740559895834,
+      "learning_rate": 0.0001,
+      "loss": 3.9716,
+      "loss/crossentropy": 2.3385006546974183,
+      "loss/hidden": 3.4359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3938352942466736,
+      "loss/reg": 0.0,
+      "step": 2020
+    },
+    {
+      "epoch": 0.013355263157894736,
+      "grad_norm": 3.15625,
+      "grad_norm_var": 0.9781158447265625,
+      "learning_rate": 0.0001,
+      "loss": 3.9531,
+      "loss/crossentropy": 2.4627522945404055,
+      "loss/hidden": 3.5578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.495425808429718,
+      "loss/reg": 0.0,
+      "step": 2030
+    },
+    {
+      "epoch": 0.013421052631578948,
+      "grad_norm": 2.75,
+      "grad_norm_var": 1.949779256184896,
+      "learning_rate": 0.0001,
+      "loss": 3.9979,
+      "loss/crossentropy": 2.11747065782547,
+      "loss/hidden": 3.59375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36948435604572294,
+      "loss/reg": 0.0,
+      "step": 2040
+    },
+    {
+      "epoch": 0.013486842105263158,
+      "grad_norm": 3.0625,
+      "grad_norm_var": 1.5490549723307292,
+      "learning_rate": 0.0001,
+      "loss": 3.8381,
+      "loss/crossentropy": 2.473706376552582,
+      "loss/hidden": 3.4265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3867632657289505,
+      "loss/reg": 0.0,
+      "step": 2050
+    },
+    {
+      "epoch": 0.013552631578947368,
+      "grad_norm": 3.125,
+      "grad_norm_var": 2.1093470786676653e+17,
+      "learning_rate": 0.0001,
+      "loss": 4.0319,
+      "loss/crossentropy": 2.4225202679634092,
+      "loss/hidden": 3.3359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3675911784172058,
+      "loss/reg": 0.0,
+      "step": 2060
+    },
+    {
+      "epoch": 0.013618421052631578,
+      "grad_norm": 5.625,
+      "grad_norm_var": 76.3380849202474,
+      "learning_rate": 0.0001,
+      "loss": 4.0746,
+      "loss/crossentropy": 2.48265061378479,
+      "loss/hidden": 3.7265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4639868468046188,
+      "loss/reg": 0.0,
+      "step": 2070
+    },
+    {
+      "epoch": 0.01368421052631579,
+      "grad_norm": 4.5,
+      "grad_norm_var": 0.7881011962890625,
+      "learning_rate": 0.0001,
+      "loss": 4.038,
+      "loss/crossentropy": 2.7775272965431212,
+      "loss/hidden": 4.0609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4948854446411133,
+      "loss/reg": 0.0,
+      "step": 2080
+    },
+    {
+      "epoch": 0.01375,
+      "grad_norm": 3.296875,
+      "grad_norm_var": 0.8111612955729167,
+      "learning_rate": 0.0001,
+      "loss": 3.8968,
+      "loss/crossentropy": 2.4613396763801574,
+      "loss/hidden": 3.3484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3684074327349663,
+      "loss/reg": 0.0,
+      "step": 2090
+    },
+    {
+      "epoch": 0.01381578947368421,
+      "grad_norm": 3.484375,
+      "grad_norm_var": 1.490550740559896,
+      "learning_rate": 0.0001,
+      "loss": 3.867,
+      "loss/crossentropy": 2.5947747588157655,
+      "loss/hidden": 3.4765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3717468947172165,
+      "loss/reg": 0.0,
+      "step": 2100
+    },
+    {
+      "epoch": 0.013881578947368422,
+      "grad_norm": 4.75,
+      "grad_norm_var": 1.7372385660807292,
+      "learning_rate": 0.0001,
+      "loss": 3.8731,
+      "loss/crossentropy": 2.466842460632324,
+      "loss/hidden": 3.2890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3417574405670166,
+      "loss/reg": 0.0,
+      "step": 2110
+    },
+    {
+      "epoch": 0.013947368421052632,
+      "grad_norm": 4.21875,
+      "grad_norm_var": 1.8861002604166666,
+      "learning_rate": 0.0001,
+      "loss": 3.823,
+      "loss/crossentropy": 2.3006282687187194,
+      "loss/hidden": 3.390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36191926896572113,
+      "loss/reg": 0.0,
+      "step": 2120
+    },
+    {
+      "epoch": 0.014013157894736842,
+      "grad_norm": 4.21875,
+      "grad_norm_var": 0.9100901285807291,
+      "learning_rate": 0.0001,
+      "loss": 3.8933,
+      "loss/crossentropy": 2.6159239768981934,
+      "loss/hidden": 3.5109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.46396631598472593,
+      "loss/reg": 0.0,
+      "step": 2130
+    },
+    {
+      "epoch": 0.014078947368421052,
+      "grad_norm": 3.734375,
+      "grad_norm_var": 0.9789388020833333,
+      "learning_rate": 0.0001,
+      "loss": 3.8761,
+      "loss/crossentropy": 2.6355370759963987,
+      "loss/hidden": 3.3421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.361694809794426,
+      "loss/reg": 0.0,
+      "step": 2140
+    },
+    {
+      "epoch": 0.014144736842105264,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.30113525390625,
+      "learning_rate": 0.0001,
+      "loss": 3.8617,
+      "loss/crossentropy": 2.6357606053352356,
+      "loss/hidden": 3.3515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3612044155597687,
+      "loss/reg": 0.0,
+      "step": 2150
+    },
+    {
+      "epoch": 0.014210526315789474,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.23430582682291667,
+      "learning_rate": 0.0001,
+      "loss": 3.8903,
+      "loss/crossentropy": 2.551873171329498,
+      "loss/hidden": 3.44375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.41060586273670197,
+      "loss/reg": 0.0,
+      "step": 2160
+    },
+    {
+      "epoch": 0.014276315789473684,
+      "grad_norm": 3.5625,
+      "grad_norm_var": 0.48121744791666665,
+      "learning_rate": 0.0001,
+      "loss": 3.9663,
+      "loss/crossentropy": 2.495719885826111,
+      "loss/hidden": 3.690625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4500987708568573,
+      "loss/reg": 0.0,
+      "step": 2170
+    },
+    {
+      "epoch": 0.014342105263157894,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.17538655598958333,
+      "learning_rate": 0.0001,
+      "loss": 3.8053,
+      "loss/crossentropy": 2.520111393928528,
+      "loss/hidden": 3.4765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.44193484634160995,
+      "loss/reg": 0.0,
+      "step": 2180
+    },
+    {
+      "epoch": 0.014407894736842106,
+      "grad_norm": 3.328125,
+      "grad_norm_var": 31.391422526041666,
+      "learning_rate": 0.0001,
+      "loss": 4.0394,
+      "loss/crossentropy": 2.5524103164672853,
+      "loss/hidden": 3.4203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38946655094623567,
+      "loss/reg": 0.0,
+      "step": 2190
+    },
+    {
+      "epoch": 0.014473684210526316,
+      "grad_norm": 8.6875,
+      "grad_norm_var": 2.0572499593098956,
+      "learning_rate": 0.0001,
+      "loss": 3.9758,
+      "loss/crossentropy": 2.2560265123844148,
+      "loss/hidden": 3.59375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4082709074020386,
+      "loss/reg": 0.0,
+      "step": 2200
+    },
+    {
+      "epoch": 0.014539473684210526,
+      "grad_norm": 3.25,
+      "grad_norm_var": 3.566722615559896,
+      "learning_rate": 0.0001,
+      "loss": 3.8174,
+      "loss/crossentropy": 2.2830474019050597,
+      "loss/hidden": 3.3375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3462225392460823,
+      "loss/reg": 0.0,
+      "step": 2210
+    },
+    {
+      "epoch": 0.014605263157894737,
+      "grad_norm": 3.109375,
+      "grad_norm_var": 0.155615234375,
+      "learning_rate": 0.0001,
+      "loss": 3.8468,
+      "loss/crossentropy": 2.3341428637504578,
+      "loss/hidden": 3.4375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4024402230978012,
+      "loss/reg": 0.0,
+      "step": 2220
+    },
+    {
+      "epoch": 0.014671052631578948,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 5.723542277018229,
+      "learning_rate": 0.0001,
+      "loss": 3.9699,
+      "loss/crossentropy": 2.201635646820068,
+      "loss/hidden": 3.2703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3397625252604485,
+      "loss/reg": 0.0,
+      "step": 2230
+    },
+    {
+      "epoch": 0.014736842105263158,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 47.54188537597656,
+      "learning_rate": 0.0001,
+      "loss": 3.9281,
+      "loss/crossentropy": 2.5724541902542115,
+      "loss/hidden": 3.2890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3728118479251862,
+      "loss/reg": 0.0,
+      "step": 2240
+    },
+    {
+      "epoch": 0.014802631578947368,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 51.28417867024739,
+      "learning_rate": 0.0001,
+      "loss": 3.8237,
+      "loss/crossentropy": 2.5087037920951842,
+      "loss/hidden": 3.2125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3448286011815071,
+      "loss/reg": 0.0,
+      "step": 2250
+    },
+    {
+      "epoch": 0.01486842105263158,
+      "grad_norm": 3.171875,
+      "grad_norm_var": 0.05579427083333333,
+      "learning_rate": 0.0001,
+      "loss": 3.8448,
+      "loss/crossentropy": 2.4813582420349123,
+      "loss/hidden": 3.409375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40958506166934966,
+      "loss/reg": 0.0,
+      "step": 2260
+    },
+    {
+      "epoch": 0.01493421052631579,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.16342671712239584,
+      "learning_rate": 0.0001,
+      "loss": 3.753,
+      "loss/crossentropy": 2.523963761329651,
+      "loss/hidden": 3.6640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38928901553153994,
+      "loss/reg": 0.0,
+      "step": 2270
+    },
+    {
+      "epoch": 0.015,
+      "grad_norm": 3.0625,
+      "grad_norm_var": 0.7111317952473958,
+      "learning_rate": 0.0001,
+      "loss": 3.8078,
+      "loss/crossentropy": 2.6787729024887086,
+      "loss/hidden": 3.71875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4244162023067474,
+      "loss/reg": 0.0,
+      "step": 2280
+    },
+    {
+      "epoch": 0.015065789473684211,
+      "grad_norm": 3.09375,
+      "grad_norm_var": 0.10339253743489583,
+      "learning_rate": 0.0001,
+      "loss": 3.7562,
+      "loss/crossentropy": 2.191652774810791,
+      "loss/hidden": 3.4140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3300579100847244,
+      "loss/reg": 0.0,
+      "step": 2290
+    },
+    {
+      "epoch": 0.015131578947368421,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.05054423014322917,
+      "learning_rate": 0.0001,
+      "loss": 3.73,
+      "loss/crossentropy": 2.4444664478302003,
+      "loss/hidden": 3.4265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4327535033226013,
+      "loss/reg": 0.0,
+      "step": 2300
+    },
+    {
+      "epoch": 0.015197368421052631,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.15128580729166666,
+      "learning_rate": 0.0001,
+      "loss": 3.7723,
+      "loss/crossentropy": 2.2329365968704225,
+      "loss/hidden": 3.51875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3683215394616127,
+      "loss/reg": 0.0,
+      "step": 2310
+    },
+    {
+      "epoch": 0.015263157894736841,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.3289713541666667,
+      "learning_rate": 0.0001,
+      "loss": 3.7699,
+      "loss/crossentropy": 2.5342983961105348,
+      "loss/hidden": 3.2921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35688025653362276,
+      "loss/reg": 0.0,
+      "step": 2320
+    },
+    {
+      "epoch": 0.015328947368421053,
+      "grad_norm": 3.15625,
+      "grad_norm_var": 0.5997060139973959,
+      "learning_rate": 0.0001,
+      "loss": 3.8494,
+      "loss/crossentropy": 2.4934002995491027,
+      "loss/hidden": 3.1546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3495032548904419,
+      "loss/reg": 0.0,
+      "step": 2330
+    },
+    {
+      "epoch": 0.015394736842105263,
+      "grad_norm": 4.28125,
+      "grad_norm_var": 1.570759073893229,
+      "learning_rate": 0.0001,
+      "loss": 3.9628,
+      "loss/crossentropy": 2.2064894437789917,
+      "loss/hidden": 3.478125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34214983880519867,
+      "loss/reg": 0.0,
+      "step": 2340
+    },
+    {
+      "epoch": 0.015460526315789473,
+      "grad_norm": 3.5,
+      "grad_norm_var": 1.9128865559895833,
+      "learning_rate": 0.0001,
+      "loss": 3.9538,
+      "loss/crossentropy": 2.5408032178878783,
+      "loss/hidden": 3.515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4111128658056259,
+      "loss/reg": 0.0,
+      "step": 2350
+    },
+    {
+      "epoch": 0.015526315789473685,
+      "grad_norm": 3.765625,
+      "grad_norm_var": 0.39661051432291666,
+      "learning_rate": 0.0001,
+      "loss": 3.8897,
+      "loss/crossentropy": 2.4922020554542543,
+      "loss/hidden": 3.3953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36137166023254397,
+      "loss/reg": 0.0,
+      "step": 2360
+    },
+    {
+      "epoch": 0.015592105263157895,
+      "grad_norm": 4.09375,
+      "grad_norm_var": 0.21868082682291667,
+      "learning_rate": 0.0001,
+      "loss": 3.8461,
+      "loss/crossentropy": 2.427833843231201,
+      "loss/hidden": 3.5921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36205882132053374,
+      "loss/reg": 0.0,
+      "step": 2370
+    },
+    {
+      "epoch": 0.015657894736842107,
+      "grad_norm": 3.171875,
+      "grad_norm_var": 0.29641520182291664,
+      "learning_rate": 0.0001,
+      "loss": 3.767,
+      "loss/crossentropy": 2.3795222878456115,
+      "loss/hidden": 3.5140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40104621052742007,
+      "loss/reg": 0.0,
+      "step": 2380
+    },
+    {
+      "epoch": 0.015723684210526317,
+      "grad_norm": 3.171875,
+      "grad_norm_var": 0.4603017171223958,
+      "learning_rate": 0.0001,
+      "loss": 3.8162,
+      "loss/crossentropy": 2.3675019264221193,
+      "loss/hidden": 3.4,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3826363369822502,
+      "loss/reg": 0.0,
+      "step": 2390
+    },
+    {
+      "epoch": 0.015789473684210527,
+      "grad_norm": 3.59375,
+      "grad_norm_var": 0.12642313639322916,
+      "learning_rate": 0.0001,
+      "loss": 3.8017,
+      "loss/crossentropy": 2.56625235080719,
+      "loss/hidden": 3.4046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3704580098390579,
+      "loss/reg": 0.0,
+      "step": 2400
+    },
+    {
+      "epoch": 0.015855263157894737,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.22603759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.8356,
+      "loss/crossentropy": 2.3642316341400145,
+      "loss/hidden": 3.640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4971610188484192,
+      "loss/reg": 0.0,
+      "step": 2410
+    },
+    {
+      "epoch": 0.015921052631578947,
+      "grad_norm": 3.359375,
+      "grad_norm_var": 2.7929354478016266e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.9524,
+      "loss/crossentropy": 2.5260600686073302,
+      "loss/hidden": 3.4734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34504298865795135,
+      "loss/reg": 0.0,
+      "step": 2420
+    },
+    {
+      "epoch": 0.015986842105263157,
+      "grad_norm": 3.140625,
+      "grad_norm_var": 2.792935447600693e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.7506,
+      "loss/crossentropy": 2.6639176845550536,
+      "loss/hidden": 3.184375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3401388913393021,
+      "loss/reg": 0.0,
+      "step": 2430
+    },
+    {
+      "epoch": 0.016052631578947367,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.0854644775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.7687,
+      "loss/crossentropy": 2.3805726766586304,
+      "loss/hidden": 3.1671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3257554292678833,
+      "loss/reg": 0.0,
+      "step": 2440
+    },
+    {
+      "epoch": 0.01611842105263158,
+      "grad_norm": 8.125,
+      "grad_norm_var": 1.995349713361273e+17,
+      "learning_rate": 0.0001,
+      "loss": 4.0533,
+      "loss/crossentropy": 2.370633268356323,
+      "loss/hidden": 3.4765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3756751254200935,
+      "loss/reg": 0.0,
+      "step": 2450
+    },
+    {
+      "epoch": 0.01618421052631579,
+      "grad_norm": 2.75,
+      "grad_norm_var": 1.995349712714498e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.7822,
+      "loss/crossentropy": 2.5394015312194824,
+      "loss/hidden": 3.3125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34547194838523865,
+      "loss/reg": 0.0,
+      "step": 2460
+    },
+    {
+      "epoch": 0.01625,
+      "grad_norm": 3.15625,
+      "grad_norm_var": 0.0999664306640625,
+      "learning_rate": 0.0001,
+      "loss": 3.6838,
+      "loss/crossentropy": 2.478586256504059,
+      "loss/hidden": 3.4109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4159191906452179,
+      "loss/reg": 0.0,
+      "step": 2470
+    },
+    {
+      "epoch": 0.01631578947368421,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.48103739420572916,
+      "learning_rate": 0.0001,
+      "loss": 3.7274,
+      "loss/crossentropy": 2.3859502553939818,
+      "loss/hidden": 3.4234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36551299393177034,
+      "loss/reg": 0.0,
+      "step": 2480
+    },
+    {
+      "epoch": 0.01638157894736842,
+      "grad_norm": 3.171875,
+      "grad_norm_var": 0.33056640625,
+      "learning_rate": 0.0001,
+      "loss": 3.8413,
+      "loss/crossentropy": 2.47422776222229,
+      "loss/hidden": 3.2703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34280748963356017,
+      "loss/reg": 0.0,
+      "step": 2490
+    },
+    {
+      "epoch": 0.01644736842105263,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.30504150390625,
+      "learning_rate": 0.0001,
+      "loss": 3.7262,
+      "loss/crossentropy": 2.288319444656372,
+      "loss/hidden": 3.2109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31078503280878067,
+      "loss/reg": 0.0,
+      "step": 2500
+    },
+    {
+      "epoch": 0.01651315789473684,
+      "grad_norm": 3.125,
+      "grad_norm_var": 0.2894683837890625,
+      "learning_rate": 0.0001,
+      "loss": 3.8194,
+      "loss/crossentropy": 2.2243799686431887,
+      "loss/hidden": 3.434375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33974049538373946,
+      "loss/reg": 0.0,
+      "step": 2510
+    },
+    {
+      "epoch": 0.016578947368421054,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.3177235921223958,
+      "learning_rate": 0.0001,
+      "loss": 3.6925,
+      "loss/crossentropy": 2.63401620388031,
+      "loss/hidden": 3.39375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.39840718507766726,
+      "loss/reg": 0.0,
+      "step": 2520
+    },
+    {
+      "epoch": 0.016644736842105264,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.2181549072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.7573,
+      "loss/crossentropy": 2.4925423860549927,
+      "loss/hidden": 3.3359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3407833933830261,
+      "loss/reg": 0.0,
+      "step": 2530
+    },
+    {
+      "epoch": 0.016710526315789474,
+      "grad_norm": 3.328125,
+      "grad_norm_var": 0.19260152180989584,
+      "learning_rate": 0.0001,
+      "loss": 3.7514,
+      "loss/crossentropy": 2.3376643657684326,
+      "loss/hidden": 3.4609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4351751744747162,
+      "loss/reg": 0.0,
+      "step": 2540
+    },
+    {
+      "epoch": 0.016776315789473684,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.12961324055989584,
+      "learning_rate": 0.0001,
+      "loss": 3.7217,
+      "loss/crossentropy": 2.279516875743866,
+      "loss/hidden": 3.58125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37157190442085264,
+      "loss/reg": 0.0,
+      "step": 2550
+    },
+    {
+      "epoch": 0.016842105263157894,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.0912017822265625,
+      "learning_rate": 0.0001,
+      "loss": 3.6839,
+      "loss/crossentropy": 2.2843039661645888,
+      "loss/hidden": 3.2125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33208019435405733,
+      "loss/reg": 0.0,
+      "step": 2560
+    },
+    {
+      "epoch": 0.016907894736842104,
+      "grad_norm": 3.28125,
+      "grad_norm_var": 0.32079671223958334,
+      "learning_rate": 0.0001,
+      "loss": 3.7381,
+      "loss/crossentropy": 2.2308380246162414,
+      "loss/hidden": 3.29375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30514844954013826,
+      "loss/reg": 0.0,
+      "step": 2570
+    },
+    {
+      "epoch": 0.016973684210526314,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.6193522135416667,
+      "learning_rate": 0.0001,
+      "loss": 3.7209,
+      "loss/crossentropy": 2.500720489025116,
+      "loss/hidden": 3.2859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38955146372318267,
+      "loss/reg": 0.0,
+      "step": 2580
+    },
+    {
+      "epoch": 0.017039473684210528,
+      "grad_norm": 3.0,
+      "grad_norm_var": 0.09539388020833334,
+      "learning_rate": 0.0001,
+      "loss": 3.6933,
+      "loss/crossentropy": 2.397514319419861,
+      "loss/hidden": 3.4796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.426153627038002,
+      "loss/reg": 0.0,
+      "step": 2590
+    },
+    {
+      "epoch": 0.017105263157894738,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 0.44207356770833334,
+      "learning_rate": 0.0001,
+      "loss": 3.7927,
+      "loss/crossentropy": 2.4746341586112974,
+      "loss/hidden": 3.4515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3806774616241455,
+      "loss/reg": 0.0,
+      "step": 2600
+    },
+    {
+      "epoch": 0.017171052631578948,
+      "grad_norm": 3.109375,
+      "grad_norm_var": 0.10927632649739584,
+      "learning_rate": 0.0001,
+      "loss": 3.7507,
+      "loss/crossentropy": 2.6908259630203246,
+      "loss/hidden": 3.459375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.47532927691936494,
+      "loss/reg": 0.0,
+      "step": 2610
+    },
+    {
+      "epoch": 0.017236842105263158,
+      "grad_norm": 2.75,
+      "grad_norm_var": 1.8751780192057292,
+      "learning_rate": 0.0001,
+      "loss": 3.8498,
+      "loss/crossentropy": 2.3184617161750793,
+      "loss/hidden": 3.3125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3832893192768097,
+      "loss/reg": 0.0,
+      "step": 2620
+    },
+    {
+      "epoch": 0.017302631578947368,
+      "grad_norm": 3.734375,
+      "grad_norm_var": 1.95693359375,
+      "learning_rate": 0.0001,
+      "loss": 3.8697,
+      "loss/crossentropy": 2.386726236343384,
+      "loss/hidden": 3.38125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4112528935074806,
+      "loss/reg": 0.0,
+      "step": 2630
+    },
+    {
+      "epoch": 0.017368421052631578,
+      "grad_norm": 3.09375,
+      "grad_norm_var": 236.78067118326823,
+      "learning_rate": 0.0001,
+      "loss": 3.843,
+      "loss/crossentropy": 2.6268559217453005,
+      "loss/hidden": 3.525,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5286044746637344,
+      "loss/reg": 0.0,
+      "step": 2640
+    },
+    {
+      "epoch": 0.017434210526315788,
+      "grad_norm": 3.15625,
+      "grad_norm_var": 238.80460510253906,
+      "learning_rate": 0.0001,
+      "loss": 3.7934,
+      "loss/crossentropy": 2.621377694606781,
+      "loss/hidden": 3.346875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.39670759439468384,
+      "loss/reg": 0.0,
+      "step": 2650
+    },
+    {
+      "epoch": 0.0175,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.4898274739583333,
+      "learning_rate": 0.0001,
+      "loss": 3.8686,
+      "loss/crossentropy": 2.3771218061447144,
+      "loss/hidden": 3.5328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.39255764335393906,
+      "loss/reg": 0.0,
+      "step": 2660
+    },
+    {
+      "epoch": 0.01756578947368421,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.9812001546223958,
+      "learning_rate": 0.0001,
+      "loss": 3.6636,
+      "loss/crossentropy": 2.2852025091648103,
+      "loss/hidden": 3.3359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3551890656352043,
+      "loss/reg": 0.0,
+      "step": 2670
+    },
+    {
+      "epoch": 0.017631578947368422,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.38752848307291665,
+      "learning_rate": 0.0001,
+      "loss": 3.7338,
+      "loss/crossentropy": 2.280995038151741,
+      "loss/hidden": 3.271875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34781029969453814,
+      "loss/reg": 0.0,
+      "step": 2680
+    },
+    {
+      "epoch": 0.017697368421052632,
+      "grad_norm": 2.953125,
+      "grad_norm_var": 0.5138661702473958,
+      "learning_rate": 0.0001,
+      "loss": 3.7751,
+      "loss/crossentropy": 2.5928542375564576,
+      "loss/hidden": 3.2203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3643882930278778,
+      "loss/reg": 0.0,
+      "step": 2690
+    },
+    {
+      "epoch": 0.017763157894736842,
+      "grad_norm": 3.015625,
+      "grad_norm_var": 2.5011057535807293,
+      "learning_rate": 0.0001,
+      "loss": 3.7281,
+      "loss/crossentropy": 2.71818265914917,
+      "loss/hidden": 3.3203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.41874536871910095,
+      "loss/reg": 0.0,
+      "step": 2700
+    },
+    {
+      "epoch": 0.017828947368421052,
+      "grad_norm": 3.0625,
+      "grad_norm_var": 0.2831939697265625,
+      "learning_rate": 0.0001,
+      "loss": 3.7723,
+      "loss/crossentropy": 2.3557824969291685,
+      "loss/hidden": 3.303125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36439308822155,
+      "loss/reg": 0.0,
+      "step": 2710
+    },
+    {
+      "epoch": 0.017894736842105262,
+      "grad_norm": 3.390625,
+      "grad_norm_var": 1.9459625244140626,
+      "learning_rate": 0.0001,
+      "loss": 3.854,
+      "loss/crossentropy": 2.3793618083000183,
+      "loss/hidden": 3.371875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36435145139694214,
+      "loss/reg": 0.0,
+      "step": 2720
+    },
+    {
+      "epoch": 0.017960526315789475,
+      "grad_norm": 2.5,
+      "grad_norm_var": 1.9445271809895834,
+      "learning_rate": 0.0001,
+      "loss": 3.7925,
+      "loss/crossentropy": 2.364825797080994,
+      "loss/hidden": 3.2453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3577578902244568,
+      "loss/reg": 0.0,
+      "step": 2730
+    },
+    {
+      "epoch": 0.018026315789473685,
+      "grad_norm": 2.9375,
+      "grad_norm_var": 0.18502197265625,
+      "learning_rate": 0.0001,
+      "loss": 3.747,
+      "loss/crossentropy": 2.3110872566699983,
+      "loss/hidden": 3.4078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3662073493003845,
+      "loss/reg": 0.0,
+      "step": 2740
+    },
+    {
+      "epoch": 0.018092105263157895,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.06238606770833333,
+      "learning_rate": 0.0001,
+      "loss": 3.736,
+      "loss/crossentropy": 2.4977880120277405,
+      "loss/hidden": 3.3578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40198240578174593,
+      "loss/reg": 0.0,
+      "step": 2750
+    },
+    {
+      "epoch": 0.018157894736842106,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 1.3276357014973958,
+      "learning_rate": 0.0001,
+      "loss": 3.7594,
+      "loss/crossentropy": 2.3260527729988096,
+      "loss/hidden": 3.39375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3752635881304741,
+      "loss/reg": 0.0,
+      "step": 2760
+    },
+    {
+      "epoch": 0.018223684210526316,
+      "grad_norm": 3.0,
+      "grad_norm_var": 0.06177978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.7632,
+      "loss/crossentropy": 2.6722695350646974,
+      "loss/hidden": 3.34375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36579819619655607,
+      "loss/reg": 0.0,
+      "step": 2770
+    },
+    {
+      "epoch": 0.018289473684210526,
+      "grad_norm": 3.78125,
+      "grad_norm_var": 0.07119038899739584,
+      "learning_rate": 0.0001,
+      "loss": 3.7821,
+      "loss/crossentropy": 2.712409424781799,
+      "loss/hidden": 3.59375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.43594706654548643,
+      "loss/reg": 0.0,
+      "step": 2780
+    },
+    {
+      "epoch": 0.018355263157894736,
+      "grad_norm": 3.21875,
+      "grad_norm_var": 2.3453409830729166,
+      "learning_rate": 0.0001,
+      "loss": 3.7973,
+      "loss/crossentropy": 2.547107517719269,
+      "loss/hidden": 3.39375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4092650800943375,
+      "loss/reg": 0.0,
+      "step": 2790
+    },
+    {
+      "epoch": 0.018421052631578946,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 4.506314086914062,
+      "learning_rate": 0.0001,
+      "loss": 3.6334,
+      "loss/crossentropy": 2.5399341940879823,
+      "loss/hidden": 3.4046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4391636699438095,
+      "loss/reg": 0.0,
+      "step": 2800
+    },
+    {
+      "epoch": 0.01848684210526316,
+      "grad_norm": 4.84375,
+      "grad_norm_var": 2.5375152587890626,
+      "learning_rate": 0.0001,
+      "loss": 3.7292,
+      "loss/crossentropy": 2.5819294929504393,
+      "loss/hidden": 3.365625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33986122310161593,
+      "loss/reg": 0.0,
+      "step": 2810
+    },
+    {
+      "epoch": 0.01855263157894737,
+      "grad_norm": 4.03125,
+      "grad_norm_var": 0.3421295166015625,
+      "learning_rate": 0.0001,
+      "loss": 3.7003,
+      "loss/crossentropy": 2.4064539194107057,
+      "loss/hidden": 3.334375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3262439340353012,
+      "loss/reg": 0.0,
+      "step": 2820
+    },
+    {
+      "epoch": 0.01861842105263158,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.1756988525390625,
+      "learning_rate": 0.0001,
+      "loss": 3.804,
+      "loss/crossentropy": 2.6151478767395018,
+      "loss/hidden": 3.6390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4677980303764343,
+      "loss/reg": 0.0,
+      "step": 2830
+    },
+    {
+      "epoch": 0.01868421052631579,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.10458577473958333,
+      "learning_rate": 0.0001,
+      "loss": 3.7672,
+      "loss/crossentropy": 2.5210029244422913,
+      "loss/hidden": 3.3953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40769249498844146,
+      "loss/reg": 0.0,
+      "step": 2840
+    },
+    {
+      "epoch": 0.01875,
+      "grad_norm": 3.28125,
+      "grad_norm_var": 0.1360504150390625,
+      "learning_rate": 0.0001,
+      "loss": 3.7829,
+      "loss/crossentropy": 2.3390053629875185,
+      "loss/hidden": 3.4421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3656760662794113,
+      "loss/reg": 0.0,
+      "step": 2850
+    },
+    {
+      "epoch": 0.01881578947368421,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.21334635416666667,
+      "learning_rate": 0.0001,
+      "loss": 3.7445,
+      "loss/crossentropy": 2.044054812192917,
+      "loss/hidden": 3.55625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.45996856689453125,
+      "loss/reg": 0.0,
+      "step": 2860
+    },
+    {
+      "epoch": 0.01888157894736842,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.35147196451822915,
+      "learning_rate": 0.0001,
+      "loss": 3.793,
+      "loss/crossentropy": 2.2911840200424196,
+      "loss/hidden": 3.5703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3373717874288559,
+      "loss/reg": 0.0,
+      "step": 2870
+    },
+    {
+      "epoch": 0.018947368421052633,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.3585845947265625,
+      "learning_rate": 0.0001,
+      "loss": 3.7057,
+      "loss/crossentropy": 2.3755346536636353,
+      "loss/hidden": 3.3609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36185318529605864,
+      "loss/reg": 0.0,
+      "step": 2880
+    },
+    {
+      "epoch": 0.019013157894736843,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.0810211181640625,
+      "learning_rate": 0.0001,
+      "loss": 3.7152,
+      "loss/crossentropy": 2.1650418758392336,
+      "loss/hidden": 3.5109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36209405958652496,
+      "loss/reg": 0.0,
+      "step": 2890
+    },
+    {
+      "epoch": 0.019078947368421053,
+      "grad_norm": 3.171875,
+      "grad_norm_var": 1.6020182291666667,
+      "learning_rate": 0.0001,
+      "loss": 3.8985,
+      "loss/crossentropy": 2.345675766468048,
+      "loss/hidden": 3.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.41470250189304353,
+      "loss/reg": 0.0,
+      "step": 2900
+    },
+    {
+      "epoch": 0.019144736842105263,
+      "grad_norm": 3.359375,
+      "grad_norm_var": 1.289264933268229,
+      "learning_rate": 0.0001,
+      "loss": 3.8444,
+      "loss/crossentropy": 2.3634544610977173,
+      "loss/hidden": 3.2984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3347387105226517,
+      "loss/reg": 0.0,
+      "step": 2910
+    },
+    {
+      "epoch": 0.019210526315789473,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.19099019368489584,
+      "learning_rate": 0.0001,
+      "loss": 3.7385,
+      "loss/crossentropy": 2.3202159285545347,
+      "loss/hidden": 3.3265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30690879598259924,
+      "loss/reg": 0.0,
+      "step": 2920
+    },
+    {
+      "epoch": 0.019276315789473683,
+      "grad_norm": 3.484375,
+      "grad_norm_var": 0.4906209309895833,
+      "learning_rate": 0.0001,
+      "loss": 3.6643,
+      "loss/crossentropy": 2.2004665434360504,
+      "loss/hidden": 3.484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3603193074464798,
+      "loss/reg": 0.0,
+      "step": 2930
+    },
+    {
+      "epoch": 0.019342105263157893,
+      "grad_norm": 3.421875,
+      "grad_norm_var": 0.13365478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.6656,
+      "loss/crossentropy": 2.5021592140197755,
+      "loss/hidden": 3.4109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35134916603565214,
+      "loss/reg": 0.0,
+      "step": 2940
+    },
+    {
+      "epoch": 0.019407894736842107,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 1.0520497639973958,
+      "learning_rate": 0.0001,
+      "loss": 3.7813,
+      "loss/crossentropy": 2.4475439548492433,
+      "loss/hidden": 3.44375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4413463234901428,
+      "loss/reg": 0.0,
+      "step": 2950
+    },
+    {
+      "epoch": 0.019473684210526317,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 1.761279296875,
+      "learning_rate": 0.0001,
+      "loss": 3.7476,
+      "loss/crossentropy": 2.57927063703537,
+      "loss/hidden": 3.4234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4805259481072426,
+      "loss/reg": 0.0,
+      "step": 2960
+    },
+    {
+      "epoch": 0.019539473684210527,
+      "grad_norm": 3.734375,
+      "grad_norm_var": 0.2982737223307292,
+      "learning_rate": 0.0001,
+      "loss": 3.7277,
+      "loss/crossentropy": 2.0291129291057586,
+      "loss/hidden": 3.0859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.285884577780962,
+      "loss/reg": 0.0,
+      "step": 2970
+    },
+    {
+      "epoch": 0.019605263157894737,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.6626942952473959,
+      "learning_rate": 0.0001,
+      "loss": 3.8596,
+      "loss/crossentropy": 2.112092435359955,
+      "loss/hidden": 3.4453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36193700730800626,
+      "loss/reg": 0.0,
+      "step": 2980
+    },
+    {
+      "epoch": 0.019671052631578947,
+      "grad_norm": 3.359375,
+      "grad_norm_var": 0.16502176920572917,
+      "learning_rate": 0.0001,
+      "loss": 3.7666,
+      "loss/crossentropy": 2.2168065547943114,
+      "loss/hidden": 3.3734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3593168243765831,
+      "loss/reg": 0.0,
+      "step": 2990
+    },
+    {
+      "epoch": 0.019736842105263157,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.45120035807291664,
+      "learning_rate": 0.0001,
+      "loss": 3.7107,
+      "loss/crossentropy": 2.350427895784378,
+      "loss/hidden": 3.3109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37478172183036806,
+      "loss/reg": 0.0,
+      "step": 3000
+    },
+    {
+      "epoch": 0.019802631578947367,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.15097249348958333,
+      "learning_rate": 0.0001,
+      "loss": 3.7559,
+      "loss/crossentropy": 2.6041540622711183,
+      "loss/hidden": 3.3421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.366491025686264,
+      "loss/reg": 0.0,
+      "step": 3010
+    },
+    {
+      "epoch": 0.01986842105263158,
+      "grad_norm": 2.953125,
+      "grad_norm_var": 0.1023834228515625,
+      "learning_rate": 0.0001,
+      "loss": 3.7322,
+      "loss/crossentropy": 2.2936912298202516,
+      "loss/hidden": 3.4234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3990582287311554,
+      "loss/reg": 0.0,
+      "step": 3020
+    },
+    {
+      "epoch": 0.01993421052631579,
+      "grad_norm": 2.9375,
+      "grad_norm_var": 0.07664388020833333,
+      "learning_rate": 0.0001,
+      "loss": 3.7405,
+      "loss/crossentropy": 2.1873862028121946,
+      "loss/hidden": 3.4703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4286995857954025,
+      "loss/reg": 0.0,
+      "step": 3030
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.3075480143229167,
+      "learning_rate": 0.0001,
+      "loss": 3.7681,
+      "loss/crossentropy": 2.0712085247039793,
+      "loss/hidden": 3.096875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2994038611650467,
+      "loss/reg": 0.0,
+      "step": 3040
+    },
+    {
+      "epoch": 0.02006578947368421,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.6994049072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.7519,
+      "loss/crossentropy": 2.155745780467987,
+      "loss/hidden": 3.2890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3106741845607758,
+      "loss/reg": 0.0,
+      "step": 3050
+    },
+    {
+      "epoch": 0.02013157894736842,
+      "grad_norm": 3.53125,
+      "grad_norm_var": 0.177978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.7618,
+      "loss/crossentropy": 2.1236796349287035,
+      "loss/hidden": 3.2046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3017591178417206,
+      "loss/reg": 0.0,
+      "step": 3060
+    },
+    {
+      "epoch": 0.02019736842105263,
+      "grad_norm": 3.265625,
+      "grad_norm_var": 0.16617431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.7225,
+      "loss/crossentropy": 2.4667672872543336,
+      "loss/hidden": 3.3328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3858541399240494,
+      "loss/reg": 0.0,
+      "step": 3070
+    },
+    {
+      "epoch": 0.02026315789473684,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 2.2265110394707968e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.8077,
+      "loss/crossentropy": 2.53361736536026,
+      "loss/hidden": 3.121875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3336446687579155,
+      "loss/reg": 0.0,
+      "step": 3080
+    },
+    {
+      "epoch": 0.020328947368421054,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 2.2265110387924992e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.7067,
+      "loss/crossentropy": 2.3802372574806214,
+      "loss/hidden": 3.2609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3561277031898499,
+      "loss/reg": 0.0,
+      "step": 3090
+    },
+    {
+      "epoch": 0.020394736842105264,
+      "grad_norm": 4.53125,
+      "grad_norm_var": 1.25572509765625,
+      "learning_rate": 0.0001,
+      "loss": 3.8816,
+      "loss/crossentropy": 2.8750504910945893,
+      "loss/hidden": 3.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37489808425307275,
+      "loss/reg": 0.0,
+      "step": 3100
+    },
+    {
+      "epoch": 0.020460526315789474,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.77646484375,
+      "learning_rate": 0.0001,
+      "loss": 3.6678,
+      "loss/crossentropy": 2.258682942390442,
+      "loss/hidden": 3.4890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3497451141476631,
+      "loss/reg": 0.0,
+      "step": 3110
+    },
+    {
+      "epoch": 0.020526315789473684,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.047412109375,
+      "learning_rate": 0.0001,
+      "loss": 3.6585,
+      "loss/crossentropy": 2.3022143959999086,
+      "loss/hidden": 3.375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37089207768440247,
+      "loss/reg": 0.0,
+      "step": 3120
+    },
+    {
+      "epoch": 0.020592105263157894,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 1.5189036051432292,
+      "learning_rate": 0.0001,
+      "loss": 3.7862,
+      "loss/crossentropy": 2.6240602493286134,
+      "loss/hidden": 3.35625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.407352888584137,
+      "loss/reg": 0.0,
+      "step": 3130
+    },
+    {
+      "epoch": 0.020657894736842104,
+      "grad_norm": 3.71875,
+      "grad_norm_var": 1.5723052978515626,
+      "learning_rate": 0.0001,
+      "loss": 3.7355,
+      "loss/crossentropy": 2.290934902429581,
+      "loss/hidden": 3.3125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3808047503232956,
+      "loss/reg": 0.0,
+      "step": 3140
+    },
+    {
+      "epoch": 0.020723684210526314,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.5337961832682292,
+      "learning_rate": 0.0001,
+      "loss": 3.7778,
+      "loss/crossentropy": 2.485193204879761,
+      "loss/hidden": 3.515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4543539136648178,
+      "loss/reg": 0.0,
+      "step": 3150
+    },
+    {
+      "epoch": 0.020789473684210528,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.3465728759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.7506,
+      "loss/crossentropy": 2.557511067390442,
+      "loss/hidden": 3.3953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4272035837173462,
+      "loss/reg": 0.0,
+      "step": 3160
+    },
+    {
+      "epoch": 0.020855263157894738,
+      "grad_norm": 3.0625,
+      "grad_norm_var": 0.20754801432291667,
+      "learning_rate": 0.0001,
+      "loss": 3.6316,
+      "loss/crossentropy": 2.310171937942505,
+      "loss/hidden": 3.184375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32435240745544436,
+      "loss/reg": 0.0,
+      "step": 3170
+    },
+    {
+      "epoch": 0.020921052631578948,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.833984375,
+      "learning_rate": 0.0001,
+      "loss": 3.7324,
+      "loss/crossentropy": 2.5518528699874876,
+      "loss/hidden": 3.3765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3816035658121109,
+      "loss/reg": 0.0,
+      "step": 3180
+    },
+    {
+      "epoch": 0.020986842105263158,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 2.3063795635792774e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.8813,
+      "loss/crossentropy": 2.3198139667510986,
+      "loss/hidden": 3.38125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3700142025947571,
+      "loss/reg": 0.0,
+      "step": 3190
+    },
+    {
+      "epoch": 0.021052631578947368,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.10243733723958333,
+      "learning_rate": 0.0001,
+      "loss": 3.6063,
+      "loss/crossentropy": 2.545992207527161,
+      "loss/hidden": 3.2875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3494896024465561,
+      "loss/reg": 0.0,
+      "step": 3200
+    },
+    {
+      "epoch": 0.021118421052631578,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.06311442057291666,
+      "learning_rate": 0.0001,
+      "loss": 3.6901,
+      "loss/crossentropy": 2.330699014663696,
+      "loss/hidden": 3.3625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34815158843994143,
+      "loss/reg": 0.0,
+      "step": 3210
+    },
+    {
+      "epoch": 0.021184210526315788,
+      "grad_norm": 3.25,
+      "grad_norm_var": 0.3830718994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.7327,
+      "loss/crossentropy": 2.489699113368988,
+      "loss/hidden": 3.325,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.337864650785923,
+      "loss/reg": 0.0,
+      "step": 3220
+    },
+    {
+      "epoch": 0.02125,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.38974609375,
+      "learning_rate": 0.0001,
+      "loss": 3.6871,
+      "loss/crossentropy": 2.3864477396011354,
+      "loss/hidden": 3.3859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4103096604347229,
+      "loss/reg": 0.0,
+      "step": 3230
+    },
+    {
+      "epoch": 0.02131578947368421,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.05084228515625,
+      "learning_rate": 0.0001,
+      "loss": 3.6458,
+      "loss/crossentropy": 2.4074989527463915,
+      "loss/hidden": 3.375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35124915838241577,
+      "loss/reg": 0.0,
+      "step": 3240
+    },
+    {
+      "epoch": 0.02138157894736842,
+      "grad_norm": 3.453125,
+      "grad_norm_var": 0.08498942057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.7083,
+      "loss/crossentropy": 2.524831974506378,
+      "loss/hidden": 3.2515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.368264502286911,
+      "loss/reg": 0.0,
+      "step": 3250
+    },
+    {
+      "epoch": 0.02144736842105263,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.42135009765625,
+      "learning_rate": 0.0001,
+      "loss": 3.7602,
+      "loss/crossentropy": 2.166905391216278,
+      "loss/hidden": 3.371875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3093524396419525,
+      "loss/reg": 0.0,
+      "step": 3260
+    },
+    {
+      "epoch": 0.02151315789473684,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.4861724853515625,
+      "learning_rate": 0.0001,
+      "loss": 3.6738,
+      "loss/crossentropy": 2.2662750601768495,
+      "loss/hidden": 3.1578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.318782140314579,
+      "loss/reg": 0.0,
+      "step": 3270
+    },
+    {
+      "epoch": 0.02157894736842105,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.13782145182291666,
+      "learning_rate": 0.0001,
+      "loss": 3.7461,
+      "loss/crossentropy": 2.0462193369865416,
+      "loss/hidden": 3.2859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30782590508461,
+      "loss/reg": 0.0,
+      "step": 3280
+    },
+    {
+      "epoch": 0.021644736842105262,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.11330973307291667,
+      "learning_rate": 0.0001,
+      "loss": 3.7008,
+      "loss/crossentropy": 2.5261476397514344,
+      "loss/hidden": 3.2734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3591727793216705,
+      "loss/reg": 0.0,
+      "step": 3290
+    },
+    {
+      "epoch": 0.021710526315789475,
+      "grad_norm": 3.0,
+      "grad_norm_var": 0.1515533447265625,
+      "learning_rate": 0.0001,
+      "loss": 3.751,
+      "loss/crossentropy": 2.569035267829895,
+      "loss/hidden": 3.3140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3920204371213913,
+      "loss/reg": 0.0,
+      "step": 3300
+    },
+    {
+      "epoch": 0.021776315789473685,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.09117431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.6385,
+      "loss/crossentropy": 2.5647154331207274,
+      "loss/hidden": 3.2078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34364444613456724,
+      "loss/reg": 0.0,
+      "step": 3310
+    },
+    {
+      "epoch": 0.021842105263157895,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 1.6411692301432292,
+      "learning_rate": 0.0001,
+      "loss": 3.6335,
+      "loss/crossentropy": 2.427185571193695,
+      "loss/hidden": 3.30625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3427447766065598,
+      "loss/reg": 0.0,
+      "step": 3320
+    },
+    {
+      "epoch": 0.021907894736842105,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.11515299479166667,
+      "learning_rate": 0.0001,
+      "loss": 3.6807,
+      "loss/crossentropy": 2.1253209471702577,
+      "loss/hidden": 3.428125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36423676908016206,
+      "loss/reg": 0.0,
+      "step": 3330
+    },
+    {
+      "epoch": 0.021973684210526315,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.046956380208333336,
+      "learning_rate": 0.0001,
+      "loss": 3.5711,
+      "loss/crossentropy": 2.3783676266670226,
+      "loss/hidden": 3.1703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3118838146328926,
+      "loss/reg": 0.0,
+      "step": 3340
+    },
+    {
+      "epoch": 0.022039473684210525,
+      "grad_norm": 2.953125,
+      "grad_norm_var": 0.060791015625,
+      "learning_rate": 0.0001,
+      "loss": 3.6252,
+      "loss/crossentropy": 2.350738251209259,
+      "loss/hidden": 3.1625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2957428440451622,
+      "loss/reg": 0.0,
+      "step": 3350
+    },
+    {
+      "epoch": 0.022105263157894735,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.09226888020833333,
+      "learning_rate": 0.0001,
+      "loss": 3.7233,
+      "loss/crossentropy": 2.5446563720703126,
+      "loss/hidden": 3.0671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3123622477054596,
+      "loss/reg": 0.0,
+      "step": 3360
+    },
+    {
+      "epoch": 0.02217105263157895,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.3047108968098958,
+      "learning_rate": 0.0001,
+      "loss": 3.6972,
+      "loss/crossentropy": 2.33814697265625,
+      "loss/hidden": 3.20625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30704180896282196,
+      "loss/reg": 0.0,
+      "step": 3370
+    },
+    {
+      "epoch": 0.02223684210526316,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.17353413899739584,
+      "learning_rate": 0.0001,
+      "loss": 3.6799,
+      "loss/crossentropy": 2.2355513691902162,
+      "loss/hidden": 3.390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3377602517604828,
+      "loss/reg": 0.0,
+      "step": 3380
+    },
+    {
+      "epoch": 0.02230263157894737,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 6.300797526041666,
+      "learning_rate": 0.0001,
+      "loss": 3.7636,
+      "loss/crossentropy": 2.3003466069698333,
+      "loss/hidden": 3.396875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3391520828008652,
+      "loss/reg": 0.0,
+      "step": 3390
+    },
+    {
+      "epoch": 0.02236842105263158,
+      "grad_norm": 3.34375,
+      "grad_norm_var": 0.0972564697265625,
+      "learning_rate": 0.0001,
+      "loss": 3.658,
+      "loss/crossentropy": 2.3254055261611937,
+      "loss/hidden": 3.246875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3125807404518127,
+      "loss/reg": 0.0,
+      "step": 3400
+    },
+    {
+      "epoch": 0.02243421052631579,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 15.84599609375,
+      "learning_rate": 0.0001,
+      "loss": 3.7692,
+      "loss/crossentropy": 2.753233790397644,
+      "loss/hidden": 3.1546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3326481133699417,
+      "loss/reg": 0.0,
+      "step": 3410
+    },
+    {
+      "epoch": 0.0225,
+      "grad_norm": 4.875,
+      "grad_norm_var": 1.10299072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.7531,
+      "loss/crossentropy": 2.282338631153107,
+      "loss/hidden": 3.3328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3607694834470749,
+      "loss/reg": 0.0,
+      "step": 3420
+    },
+    {
+      "epoch": 0.02256578947368421,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.626123046875,
+      "learning_rate": 0.0001,
+      "loss": 3.8275,
+      "loss/crossentropy": 2.5421807527542115,
+      "loss/hidden": 3.425,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5308041572570801,
+      "loss/reg": 0.0,
+      "step": 3430
+    },
+    {
+      "epoch": 0.022631578947368423,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.40051676432291666,
+      "learning_rate": 0.0001,
+      "loss": 3.7523,
+      "loss/crossentropy": 2.541818845272064,
+      "loss/hidden": 3.3625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.359403657913208,
+      "loss/reg": 0.0,
+      "step": 3440
+    },
+    {
+      "epoch": 0.022697368421052633,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.35347391764322916,
+      "learning_rate": 0.0001,
+      "loss": 3.6627,
+      "loss/crossentropy": 2.5443089246749877,
+      "loss/hidden": 3.2125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32425140738487246,
+      "loss/reg": 0.0,
+      "step": 3450
+    },
+    {
+      "epoch": 0.022763157894736843,
+      "grad_norm": 3.296875,
+      "grad_norm_var": 0.09103190104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.6541,
+      "loss/crossentropy": 2.4523619592189787,
+      "loss/hidden": 3.4453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3741248741745949,
+      "loss/reg": 0.0,
+      "step": 3460
+    },
+    {
+      "epoch": 0.022828947368421053,
+      "grad_norm": 5.375,
+      "grad_norm_var": 1.3390533447265625,
+      "learning_rate": 0.0001,
+      "loss": 3.7552,
+      "loss/crossentropy": 2.282164466381073,
+      "loss/hidden": 3.3984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32355323880910875,
+      "loss/reg": 0.0,
+      "step": 3470
+    },
+    {
+      "epoch": 0.022894736842105263,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.5425608317057292,
+      "learning_rate": 0.0001,
+      "loss": 3.6438,
+      "loss/crossentropy": 2.584494400024414,
+      "loss/hidden": 3.325,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3407262712717056,
+      "loss/reg": 0.0,
+      "step": 3480
+    },
+    {
+      "epoch": 0.022960526315789473,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 11.746613566080729,
+      "learning_rate": 0.0001,
+      "loss": 3.8186,
+      "loss/crossentropy": 2.781242084503174,
+      "loss/hidden": 3.2421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.516629433631897,
+      "loss/reg": 0.0,
+      "step": 3490
+    },
+    {
+      "epoch": 0.023026315789473683,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.38925374348958336,
+      "learning_rate": 0.0001,
+      "loss": 3.6824,
+      "loss/crossentropy": 2.6085395932197573,
+      "loss/hidden": 3.55625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40641255080699923,
+      "loss/reg": 0.0,
+      "step": 3500
+    },
+    {
+      "epoch": 0.023092105263157896,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.0653228759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.6125,
+      "loss/crossentropy": 2.143614149093628,
+      "loss/hidden": 3.2953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32446493208408356,
+      "loss/reg": 0.0,
+      "step": 3510
+    },
+    {
+      "epoch": 0.023157894736842106,
+      "grad_norm": 3.78125,
+      "grad_norm_var": 0.24351806640625,
+      "learning_rate": 0.0001,
+      "loss": 3.7227,
+      "loss/crossentropy": 2.4581753849983214,
+      "loss/hidden": 3.5296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4273629605770111,
+      "loss/reg": 0.0,
+      "step": 3520
+    },
+    {
+      "epoch": 0.023223684210526317,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.3322987874348958,
+      "learning_rate": 0.0001,
+      "loss": 3.6619,
+      "loss/crossentropy": 2.452810299396515,
+      "loss/hidden": 3.41875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.364976304769516,
+      "loss/reg": 0.0,
+      "step": 3530
+    },
+    {
+      "epoch": 0.023289473684210527,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.050780232747395834,
+      "learning_rate": 0.0001,
+      "loss": 3.5957,
+      "loss/crossentropy": 2.6891199111938477,
+      "loss/hidden": 3.0734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3124883592128754,
+      "loss/reg": 0.0,
+      "step": 3540
+    },
+    {
+      "epoch": 0.023355263157894737,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.08567301432291667,
+      "learning_rate": 0.0001,
+      "loss": 3.7773,
+      "loss/crossentropy": 2.565224659442902,
+      "loss/hidden": 3.1421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30210830420255663,
+      "loss/reg": 0.0,
+      "step": 3550
+    },
+    {
+      "epoch": 0.023421052631578947,
+      "grad_norm": 3.921875,
+      "grad_norm_var": 0.20045166015625,
+      "learning_rate": 0.0001,
+      "loss": 3.6777,
+      "loss/crossentropy": 2.3339913129806518,
+      "loss/hidden": 3.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3782587692141533,
+      "loss/reg": 0.0,
+      "step": 3560
+    },
+    {
+      "epoch": 0.023486842105263157,
+      "grad_norm": 6.375,
+      "grad_norm_var": 7.792964680989583,
+      "learning_rate": 0.0001,
+      "loss": 3.9088,
+      "loss/crossentropy": 2.305986249446869,
+      "loss/hidden": 3.2078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3260203331708908,
+      "loss/reg": 0.0,
+      "step": 3570
+    },
+    {
+      "epoch": 0.023552631578947367,
+      "grad_norm": 2.5,
+      "grad_norm_var": 1.226398722330729,
+      "learning_rate": 0.0001,
+      "loss": 3.5858,
+      "loss/crossentropy": 2.5068121433258055,
+      "loss/hidden": 3.29375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38687763810157777,
+      "loss/reg": 0.0,
+      "step": 3580
+    },
+    {
+      "epoch": 0.02361842105263158,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.08307291666666666,
+      "learning_rate": 0.0001,
+      "loss": 3.5749,
+      "loss/crossentropy": 2.440618324279785,
+      "loss/hidden": 3.246875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3219692587852478,
+      "loss/reg": 0.0,
+      "step": 3590
+    },
+    {
+      "epoch": 0.02368421052631579,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.2329742431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.7091,
+      "loss/crossentropy": 2.3230647802352906,
+      "loss/hidden": 3.4234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3276051238179207,
+      "loss/reg": 0.0,
+      "step": 3600
+    },
+    {
+      "epoch": 0.02375,
+      "grad_norm": 3.125,
+      "grad_norm_var": 0.06494038899739583,
+      "learning_rate": 0.0001,
+      "loss": 3.6957,
+      "loss/crossentropy": 2.549039614200592,
+      "loss/hidden": 3.3265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4018037021160126,
+      "loss/reg": 0.0,
+      "step": 3610
+    },
+    {
+      "epoch": 0.02381578947368421,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.05314839680989583,
+      "learning_rate": 0.0001,
+      "loss": 3.6018,
+      "loss/crossentropy": 2.5681329488754274,
+      "loss/hidden": 3.3046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3874122858047485,
+      "loss/reg": 0.0,
+      "step": 3620
+    },
+    {
+      "epoch": 0.02388157894736842,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.30606180826822915,
+      "learning_rate": 0.0001,
+      "loss": 3.6534,
+      "loss/crossentropy": 2.224066364765167,
+      "loss/hidden": 3.41875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3868778973817825,
+      "loss/reg": 0.0,
+      "step": 3630
+    },
+    {
+      "epoch": 0.02394736842105263,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.44580078125,
+      "learning_rate": 0.0001,
+      "loss": 3.7632,
+      "loss/crossentropy": 2.313184142112732,
+      "loss/hidden": 3.29375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35929109454154967,
+      "loss/reg": 0.0,
+      "step": 3640
+    },
+    {
+      "epoch": 0.02401315789473684,
+      "grad_norm": 4.65625,
+      "grad_norm_var": 0.43798421223958334,
+      "learning_rate": 0.0001,
+      "loss": 3.6503,
+      "loss/crossentropy": 2.483446490764618,
+      "loss/hidden": 3.2515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.324999064207077,
+      "loss/reg": 0.0,
+      "step": 3650
+    },
+    {
+      "epoch": 0.024078947368421054,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.271875,
+      "learning_rate": 0.0001,
+      "loss": 3.7069,
+      "loss/crossentropy": 2.2979444444179533,
+      "loss/hidden": 3.2421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3168840616941452,
+      "loss/reg": 0.0,
+      "step": 3660
+    },
+    {
+      "epoch": 0.024144736842105264,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.0767242431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.6808,
+      "loss/crossentropy": 2.4076380014419554,
+      "loss/hidden": 3.55625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3966252237558365,
+      "loss/reg": 0.0,
+      "step": 3670
+    },
+    {
+      "epoch": 0.024210526315789474,
+      "grad_norm": 3.109375,
+      "grad_norm_var": 0.06747639973958333,
+      "learning_rate": 0.0001,
+      "loss": 3.7852,
+      "loss/crossentropy": 2.6107199430465697,
+      "loss/hidden": 3.3359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3948290854692459,
+      "loss/reg": 0.0,
+      "step": 3680
+    },
+    {
+      "epoch": 0.024276315789473684,
+      "grad_norm": 3.765625,
+      "grad_norm_var": 0.13434244791666666,
+      "learning_rate": 0.0001,
+      "loss": 3.6048,
+      "loss/crossentropy": 2.5476237654685976,
+      "loss/hidden": 3.209375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.352567557990551,
+      "loss/reg": 0.0,
+      "step": 3690
+    },
+    {
+      "epoch": 0.024342105263157894,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.6126139322916667,
+      "learning_rate": 0.0001,
+      "loss": 3.7034,
+      "loss/crossentropy": 2.435033369064331,
+      "loss/hidden": 3.275,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37708690464496614,
+      "loss/reg": 0.0,
+      "step": 3700
+    },
+    {
+      "epoch": 0.024407894736842104,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.7496002197265625,
+      "learning_rate": 0.0001,
+      "loss": 3.6696,
+      "loss/crossentropy": 2.550716495513916,
+      "loss/hidden": 3.3328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38517349362373354,
+      "loss/reg": 0.0,
+      "step": 3710
+    },
+    {
+      "epoch": 0.024473684210526314,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.5516886393229167,
+      "learning_rate": 0.0001,
+      "loss": 3.5623,
+      "loss/crossentropy": 2.3806477397680283,
+      "loss/hidden": 3.153125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3013214536011219,
+      "loss/reg": 0.0,
+      "step": 3720
+    },
+    {
+      "epoch": 0.024539473684210528,
+      "grad_norm": 3.1875,
+      "grad_norm_var": 0.4890207926432292,
+      "learning_rate": 0.0001,
+      "loss": 3.6412,
+      "loss/crossentropy": 2.3859506011009217,
+      "loss/hidden": 3.428125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.49241943359375,
+      "loss/reg": 0.0,
+      "step": 3730
+    },
+    {
+      "epoch": 0.024605263157894738,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.2377838134765625,
+      "learning_rate": 0.0001,
+      "loss": 3.5895,
+      "loss/crossentropy": 2.141975212097168,
+      "loss/hidden": 3.2140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3277399495244026,
+      "loss/reg": 0.0,
+      "step": 3740
+    },
+    {
+      "epoch": 0.024671052631578948,
+      "grad_norm": 3.140625,
+      "grad_norm_var": 0.12506510416666666,
+      "learning_rate": 0.0001,
+      "loss": 3.6233,
+      "loss/crossentropy": 2.4696611404418944,
+      "loss/hidden": 3.3265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3837138593196869,
+      "loss/reg": 0.0,
+      "step": 3750
+    },
+    {
+      "epoch": 0.024736842105263158,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 2.6720540364583334,
+      "learning_rate": 0.0001,
+      "loss": 3.7719,
+      "loss/crossentropy": 2.494647514820099,
+      "loss/hidden": 3.275,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37584047913551333,
+      "loss/reg": 0.0,
+      "step": 3760
+    },
+    {
+      "epoch": 0.024802631578947368,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.63804931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.676,
+      "loss/crossentropy": 2.5242549180984497,
+      "loss/hidden": 3.2328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.39244888722896576,
+      "loss/reg": 0.0,
+      "step": 3770
+    },
+    {
+      "epoch": 0.024868421052631578,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.3232167561848958,
+      "learning_rate": 0.0001,
+      "loss": 3.5206,
+      "loss/crossentropy": 2.2167584180831907,
+      "loss/hidden": 3.109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31152922809123995,
+      "loss/reg": 0.0,
+      "step": 3780
+    },
+    {
+      "epoch": 0.024934210526315788,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.8850331624348958,
+      "learning_rate": 0.0001,
+      "loss": 3.6831,
+      "loss/crossentropy": 2.406609225273132,
+      "loss/hidden": 3.2734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3529895097017288,
+      "loss/reg": 0.0,
+      "step": 3790
+    },
+    {
+      "epoch": 0.025,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.8739491780598958,
+      "learning_rate": 0.0001,
+      "loss": 3.6024,
+      "loss/crossentropy": 2.270749258995056,
+      "loss/hidden": 3.3390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32701381742954255,
+      "loss/reg": 0.0,
+      "step": 3800
+    },
+    {
+      "epoch": 0.02506578947368421,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.06500651041666666,
+      "learning_rate": 0.0001,
+      "loss": 3.6037,
+      "loss/crossentropy": 2.3936703205108643,
+      "loss/hidden": 3.3078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4090299874544144,
+      "loss/reg": 0.0,
+      "step": 3810
+    },
+    {
+      "epoch": 0.02513157894736842,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.3386301676432292,
+      "learning_rate": 0.0001,
+      "loss": 3.8203,
+      "loss/crossentropy": 2.057803177833557,
+      "loss/hidden": 3.390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3018879994750023,
+      "loss/reg": 0.0,
+      "step": 3820
+    },
+    {
+      "epoch": 0.02519736842105263,
+      "grad_norm": 10.375,
+      "grad_norm_var": 3.7443756103515624,
+      "learning_rate": 0.0001,
+      "loss": 3.5987,
+      "loss/crossentropy": 2.558328187465668,
+      "loss/hidden": 3.3671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40375421941280365,
+      "loss/reg": 0.0,
+      "step": 3830
+    },
+    {
+      "epoch": 0.02526315789473684,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 4.824331665039063,
+      "learning_rate": 0.0001,
+      "loss": 3.6109,
+      "loss/crossentropy": 2.4123119592666624,
+      "loss/hidden": 3.1921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3224829614162445,
+      "loss/reg": 0.0,
+      "step": 3840
+    },
+    {
+      "epoch": 0.02532894736842105,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 1.4220611572265625,
+      "learning_rate": 0.0001,
+      "loss": 3.6503,
+      "loss/crossentropy": 2.4460156679153444,
+      "loss/hidden": 3.3140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37439659237861633,
+      "loss/reg": 0.0,
+      "step": 3850
+    },
+    {
+      "epoch": 0.02539473684210526,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.24488525390625,
+      "learning_rate": 0.0001,
+      "loss": 3.6799,
+      "loss/crossentropy": 2.471416544914246,
+      "loss/hidden": 3.265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3233081191778183,
+      "loss/reg": 0.0,
+      "step": 3860
+    },
+    {
+      "epoch": 0.025460526315789475,
+      "grad_norm": 4.6875,
+      "grad_norm_var": 2.005939737955729,
+      "learning_rate": 0.0001,
+      "loss": 3.7099,
+      "loss/crossentropy": 2.46109699010849,
+      "loss/hidden": 3.178125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31833461821079256,
+      "loss/reg": 0.0,
+      "step": 3870
+    },
+    {
+      "epoch": 0.025526315789473685,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.4576568603515625,
+      "learning_rate": 0.0001,
+      "loss": 3.62,
+      "loss/crossentropy": 2.4385437607765197,
+      "loss/hidden": 3.1953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3444881528615952,
+      "loss/reg": 0.0,
+      "step": 3880
+    },
+    {
+      "epoch": 0.025592105263157895,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.07125244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.5814,
+      "loss/crossentropy": 2.464188981056213,
+      "loss/hidden": 3.2640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4162255361676216,
+      "loss/reg": 0.0,
+      "step": 3890
+    },
+    {
+      "epoch": 0.025657894736842105,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.13810221354166666,
+      "learning_rate": 0.0001,
+      "loss": 3.5963,
+      "loss/crossentropy": 2.5776121497154234,
+      "loss/hidden": 3.1296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.351967790722847,
+      "loss/reg": 0.0,
+      "step": 3900
+    },
+    {
+      "epoch": 0.025723684210526315,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.13177083333333334,
+      "learning_rate": 0.0001,
+      "loss": 3.6137,
+      "loss/crossentropy": 2.4709773540496824,
+      "loss/hidden": 3.28125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3811213612556458,
+      "loss/reg": 0.0,
+      "step": 3910
+    },
+    {
+      "epoch": 0.025789473684210525,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.047591145833333334,
+      "learning_rate": 0.0001,
+      "loss": 3.5243,
+      "loss/crossentropy": 2.3289324045181274,
+      "loss/hidden": 3.240625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32472735941410064,
+      "loss/reg": 0.0,
+      "step": 3920
+    },
+    {
+      "epoch": 0.025855263157894735,
+      "grad_norm": 3.359375,
+      "grad_norm_var": 0.07144266764322917,
+      "learning_rate": 0.0001,
+      "loss": 3.5618,
+      "loss/crossentropy": 2.151239442825317,
+      "loss/hidden": 3.15,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3790448889136314,
+      "loss/reg": 0.0,
+      "step": 3930
+    },
+    {
+      "epoch": 0.02592105263157895,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.19239908854166668,
+      "learning_rate": 0.0001,
+      "loss": 3.5689,
+      "loss/crossentropy": 2.281427323818207,
+      "loss/hidden": 3.321875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32773717790842055,
+      "loss/reg": 0.0,
+      "step": 3940
+    },
+    {
+      "epoch": 0.02598684210526316,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.07558186848958333,
+      "learning_rate": 0.0001,
+      "loss": 3.5047,
+      "loss/crossentropy": 2.2366716623306275,
+      "loss/hidden": 3.36875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3462549954652786,
+      "loss/reg": 0.0,
+      "step": 3950
+    },
+    {
+      "epoch": 0.02605263157894737,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.0950347900390625,
+      "learning_rate": 0.0001,
+      "loss": 3.6604,
+      "loss/crossentropy": 2.6565569043159485,
+      "loss/hidden": 3.3375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38759642243385317,
+      "loss/reg": 0.0,
+      "step": 3960
+    },
+    {
+      "epoch": 0.02611842105263158,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.04934895833333333,
+      "learning_rate": 0.0001,
+      "loss": 3.4984,
+      "loss/crossentropy": 2.3093223094940187,
+      "loss/hidden": 3.2484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3547346442937851,
+      "loss/reg": 0.0,
+      "step": 3970
+    },
+    {
+      "epoch": 0.02618421052631579,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.6375284830729167,
+      "learning_rate": 0.0001,
+      "loss": 3.6552,
+      "loss/crossentropy": 2.5669564962387086,
+      "loss/hidden": 3.640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3741306886076927,
+      "loss/reg": 0.0,
+      "step": 3980
+    },
+    {
+      "epoch": 0.02625,
+      "grad_norm": 3.328125,
+      "grad_norm_var": 0.1037994384765625,
+      "learning_rate": 0.0001,
+      "loss": 3.707,
+      "loss/crossentropy": 2.255212366580963,
+      "loss/hidden": 3.4078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36638626754283904,
+      "loss/reg": 0.0,
+      "step": 3990
+    },
+    {
+      "epoch": 0.02631578947368421,
+      "grad_norm": 3.484375,
+      "grad_norm_var": 0.1186431884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.5907,
+      "loss/crossentropy": 2.5717132806777956,
+      "loss/hidden": 3.159375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3674279361963272,
+      "loss/reg": 0.0,
+      "step": 4000
+    },
+    {
+      "epoch": 0.026381578947368423,
+      "grad_norm": 3.09375,
+      "grad_norm_var": 0.09750874837239583,
+      "learning_rate": 0.0001,
+      "loss": 3.5937,
+      "loss/crossentropy": 2.494677722454071,
+      "loss/hidden": 3.209375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34122500121593474,
+      "loss/reg": 0.0,
+      "step": 4010
+    },
+    {
+      "epoch": 0.026447368421052633,
+      "grad_norm": 2.9375,
+      "grad_norm_var": 0.07803446451822917,
+      "learning_rate": 0.0001,
+      "loss": 3.4996,
+      "loss/crossentropy": 2.4037094593048094,
+      "loss/hidden": 3.1328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34369616210460663,
+      "loss/reg": 0.0,
+      "step": 4020
+    },
+    {
+      "epoch": 0.026513157894736843,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.031087239583333332,
+      "learning_rate": 0.0001,
+      "loss": 3.5883,
+      "loss/crossentropy": 2.5161670804023744,
+      "loss/hidden": 3.234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3320692449808121,
+      "loss/reg": 0.0,
+      "step": 4030
+    },
+    {
+      "epoch": 0.026578947368421053,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.27988993326822914,
+      "learning_rate": 0.0001,
+      "loss": 3.5987,
+      "loss/crossentropy": 2.489400029182434,
+      "loss/hidden": 3.2828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37186973094940184,
+      "loss/reg": 0.0,
+      "step": 4040
+    },
+    {
+      "epoch": 0.026644736842105263,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.2902740478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.6423,
+      "loss/crossentropy": 2.1810465335845945,
+      "loss/hidden": 3.140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30852093994617463,
+      "loss/reg": 0.0,
+      "step": 4050
+    },
+    {
+      "epoch": 0.026710526315789473,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.2831776936848958,
+      "learning_rate": 0.0001,
+      "loss": 3.5696,
+      "loss/crossentropy": 2.5600404262542726,
+      "loss/hidden": 3.2265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38900414407253264,
+      "loss/reg": 0.0,
+      "step": 4060
+    },
+    {
+      "epoch": 0.026776315789473683,
+      "grad_norm": 4.125,
+      "grad_norm_var": 0.25816141764322914,
+      "learning_rate": 0.0001,
+      "loss": 3.6429,
+      "loss/crossentropy": 2.4915748476982116,
+      "loss/hidden": 3.290625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35222682952880857,
+      "loss/reg": 0.0,
+      "step": 4070
+    },
+    {
+      "epoch": 0.026842105263157896,
+      "grad_norm": 3.875,
+      "grad_norm_var": 2.97427978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.6704,
+      "loss/crossentropy": 2.131973624229431,
+      "loss/hidden": 3.36875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32549644112586973,
+      "loss/reg": 0.0,
+      "step": 4080
+    },
+    {
+      "epoch": 0.026907894736842106,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.25222981770833336,
+      "learning_rate": 0.0001,
+      "loss": 3.591,
+      "loss/crossentropy": 2.196081441640854,
+      "loss/hidden": 3.240625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35593045353889463,
+      "loss/reg": 0.0,
+      "step": 4090
+    },
+    {
+      "epoch": 0.026973684210526316,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.20009765625,
+      "learning_rate": 0.0001,
+      "loss": 3.5044,
+      "loss/crossentropy": 2.3047094464302065,
+      "loss/hidden": 3.159375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2999614104628563,
+      "loss/reg": 0.0,
+      "step": 4100
+    },
+    {
+      "epoch": 0.027039473684210526,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.5296051025390625,
+      "learning_rate": 0.0001,
+      "loss": 3.6015,
+      "loss/crossentropy": 2.4926783800125123,
+      "loss/hidden": 3.2578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.342082779109478,
+      "loss/reg": 0.0,
+      "step": 4110
+    },
+    {
+      "epoch": 0.027105263157894736,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.05623270670572917,
+      "learning_rate": 0.0001,
+      "loss": 3.5623,
+      "loss/crossentropy": 2.6063008666038514,
+      "loss/hidden": 3.128125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3086448922753334,
+      "loss/reg": 0.0,
+      "step": 4120
+    },
+    {
+      "epoch": 0.027171052631578946,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.103515625,
+      "learning_rate": 0.0001,
+      "loss": 3.5628,
+      "loss/crossentropy": 2.516204798221588,
+      "loss/hidden": 3.2609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36054509580135347,
+      "loss/reg": 0.0,
+      "step": 4130
+    },
+    {
+      "epoch": 0.027236842105263157,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.09970601399739583,
+      "learning_rate": 0.0001,
+      "loss": 3.5328,
+      "loss/crossentropy": 2.5081961393356322,
+      "loss/hidden": 3.2828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3743078649044037,
+      "loss/reg": 0.0,
+      "step": 4140
+    },
+    {
+      "epoch": 0.02730263157894737,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.0626373291015625,
+      "learning_rate": 0.0001,
+      "loss": 3.5896,
+      "loss/crossentropy": 2.386087703704834,
+      "loss/hidden": 3.225,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33356338143348696,
+      "loss/reg": 0.0,
+      "step": 4150
+    },
+    {
+      "epoch": 0.02736842105263158,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.12830301920572917,
+      "learning_rate": 0.0001,
+      "loss": 3.6293,
+      "loss/crossentropy": 2.2993146777153015,
+      "loss/hidden": 3.134375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28893803358078,
+      "loss/reg": 0.0,
+      "step": 4160
+    },
+    {
+      "epoch": 0.02743421052631579,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.29797770182291666,
+      "learning_rate": 0.0001,
+      "loss": 3.5201,
+      "loss/crossentropy": 2.4818823099136353,
+      "loss/hidden": 3.309375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3620707929134369,
+      "loss/reg": 0.0,
+      "step": 4170
+    },
+    {
+      "epoch": 0.0275,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.2361328125,
+      "learning_rate": 0.0001,
+      "loss": 3.533,
+      "loss/crossentropy": 2.4130859971046448,
+      "loss/hidden": 3.10625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3289807617664337,
+      "loss/reg": 0.0,
+      "step": 4180
+    },
+    {
+      "epoch": 0.02756578947368421,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.06122945149739583,
+      "learning_rate": 0.0001,
+      "loss": 3.4961,
+      "loss/crossentropy": 2.3559444665908815,
+      "loss/hidden": 3.1140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3270682215690613,
+      "loss/reg": 0.0,
+      "step": 4190
+    },
+    {
+      "epoch": 0.02763157894736842,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.3246378580729167,
+      "learning_rate": 0.0001,
+      "loss": 3.7095,
+      "loss/crossentropy": 2.370071732997894,
+      "loss/hidden": 3.1328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36643182039260863,
+      "loss/reg": 0.0,
+      "step": 4200
+    },
+    {
+      "epoch": 0.02769736842105263,
+      "grad_norm": 8.375,
+      "grad_norm_var": 2.272069295247396,
+      "learning_rate": 0.0001,
+      "loss": 3.5404,
+      "loss/crossentropy": 2.4906920313835146,
+      "loss/hidden": 3.2359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3274578660726547,
+      "loss/reg": 0.0,
+      "step": 4210
+    },
+    {
+      "epoch": 0.027763157894736844,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 2.19049072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.6801,
+      "loss/crossentropy": 2.66324725151062,
+      "loss/hidden": 3.5046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38620950281620026,
+      "loss/reg": 0.0,
+      "step": 4220
+    },
+    {
+      "epoch": 0.027828947368421054,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.1225494384765625,
+      "learning_rate": 0.0001,
+      "loss": 3.6293,
+      "loss/crossentropy": 2.3201419711112976,
+      "loss/hidden": 3.284375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33118238747119905,
+      "loss/reg": 0.0,
+      "step": 4230
+    },
+    {
+      "epoch": 0.027894736842105264,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.062108357747395836,
+      "learning_rate": 0.0001,
+      "loss": 3.5523,
+      "loss/crossentropy": 2.2000674962997437,
+      "loss/hidden": 3.1953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33759562224149703,
+      "loss/reg": 0.0,
+      "step": 4240
+    },
+    {
+      "epoch": 0.027960526315789474,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.10436197916666666,
+      "learning_rate": 0.0001,
+      "loss": 3.6554,
+      "loss/crossentropy": 2.4001947045326233,
+      "loss/hidden": 3.3578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34554801881313324,
+      "loss/reg": 0.0,
+      "step": 4250
+    },
+    {
+      "epoch": 0.028026315789473684,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.09031473795572917,
+      "learning_rate": 0.0001,
+      "loss": 3.5922,
+      "loss/crossentropy": 2.5996686697006224,
+      "loss/hidden": 3.209375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34134136140346527,
+      "loss/reg": 0.0,
+      "step": 4260
+    },
+    {
+      "epoch": 0.028092105263157894,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.07454020182291667,
+      "learning_rate": 0.0001,
+      "loss": 3.5895,
+      "loss/crossentropy": 2.1346129894256594,
+      "loss/hidden": 3.1671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2764407262206078,
+      "loss/reg": 0.0,
+      "step": 4270
+    },
+    {
+      "epoch": 0.028157894736842104,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.07390950520833334,
+      "learning_rate": 0.0001,
+      "loss": 3.6094,
+      "loss/crossentropy": 2.539172089099884,
+      "loss/hidden": 3.3578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3945852980017662,
+      "loss/reg": 0.0,
+      "step": 4280
+    },
+    {
+      "epoch": 0.028223684210526317,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.14036051432291666,
+      "learning_rate": 0.0001,
+      "loss": 3.6505,
+      "loss/crossentropy": 2.4181793212890623,
+      "loss/hidden": 3.484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.45307959616184235,
+      "loss/reg": 0.0,
+      "step": 4290
+    },
+    {
+      "epoch": 0.028289473684210528,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.14678446451822916,
+      "learning_rate": 0.0001,
+      "loss": 3.5539,
+      "loss/crossentropy": 2.3630972266197205,
+      "loss/hidden": 3.0015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28445124477148054,
+      "loss/reg": 0.0,
+      "step": 4300
+    },
+    {
+      "epoch": 0.028355263157894738,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.3001627604166667,
+      "learning_rate": 0.0001,
+      "loss": 3.5098,
+      "loss/crossentropy": 2.4898954033851624,
+      "loss/hidden": 3.109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.320017996430397,
+      "loss/reg": 0.0,
+      "step": 4310
+    },
+    {
+      "epoch": 0.028421052631578948,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.13015034993489583,
+      "learning_rate": 0.0001,
+      "loss": 3.5755,
+      "loss/crossentropy": 2.2955414295196532,
+      "loss/hidden": 3.434375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.41727418303489683,
+      "loss/reg": 0.0,
+      "step": 4320
+    },
+    {
+      "epoch": 0.028486842105263158,
+      "grad_norm": 3.015625,
+      "grad_norm_var": 0.29983723958333336,
+      "learning_rate": 0.0001,
+      "loss": 3.7538,
+      "loss/crossentropy": 2.4246325135231017,
+      "loss/hidden": 3.3546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34737818390131,
+      "loss/reg": 0.0,
+      "step": 4330
+    },
+    {
+      "epoch": 0.028552631578947368,
+      "grad_norm": 5.3125,
+      "grad_norm_var": 0.5324544270833333,
+      "learning_rate": 0.0001,
+      "loss": 3.6927,
+      "loss/crossentropy": 2.393894040584564,
+      "loss/hidden": 3.321875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37734392285346985,
+      "loss/reg": 0.0,
+      "step": 4340
+    },
+    {
+      "epoch": 0.028618421052631578,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.5618316650390625,
+      "learning_rate": 0.0001,
+      "loss": 3.6563,
+      "loss/crossentropy": 2.5759302139282227,
+      "loss/hidden": 3.2125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32841147780418395,
+      "loss/reg": 0.0,
+      "step": 4350
+    },
+    {
+      "epoch": 0.028684210526315788,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.06923726399739584,
+      "learning_rate": 0.0001,
+      "loss": 3.5818,
+      "loss/crossentropy": 2.664570915699005,
+      "loss/hidden": 3.040625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31884663701057436,
+      "loss/reg": 0.0,
+      "step": 4360
+    },
+    {
+      "epoch": 0.02875,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.219873046875,
+      "learning_rate": 0.0001,
+      "loss": 3.6444,
+      "loss/crossentropy": 2.3797228574752807,
+      "loss/hidden": 3.5796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4052841871976852,
+      "loss/reg": 0.0,
+      "step": 4370
+    },
+    {
+      "epoch": 0.02881578947368421,
+      "grad_norm": 3.0,
+      "grad_norm_var": 0.5880849202473958,
+      "learning_rate": 0.0001,
+      "loss": 3.5881,
+      "loss/crossentropy": 2.5849244236946105,
+      "loss/hidden": 3.5046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.529149529337883,
+      "loss/reg": 0.0,
+      "step": 4380
+    },
+    {
+      "epoch": 0.02888157894736842,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.31160380045572916,
+      "learning_rate": 0.0001,
+      "loss": 3.4738,
+      "loss/crossentropy": 2.5852147936820984,
+      "loss/hidden": 3.1875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3336706295609474,
+      "loss/reg": 0.0,
+      "step": 4390
+    },
+    {
+      "epoch": 0.02894736842105263,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.05146077473958333,
+      "learning_rate": 0.0001,
+      "loss": 3.6199,
+      "loss/crossentropy": 2.412027895450592,
+      "loss/hidden": 3.375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4120332598686218,
+      "loss/reg": 0.0,
+      "step": 4400
+    },
+    {
+      "epoch": 0.02901315789473684,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.138232421875,
+      "learning_rate": 0.0001,
+      "loss": 3.4929,
+      "loss/crossentropy": 2.270553803443909,
+      "loss/hidden": 3.2546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34523763358592985,
+      "loss/reg": 0.0,
+      "step": 4410
+    },
+    {
+      "epoch": 0.02907894736842105,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.1102691650390625,
+      "learning_rate": 0.0001,
+      "loss": 3.5778,
+      "loss/crossentropy": 2.361116898059845,
+      "loss/hidden": 3.246875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34047031700611113,
+      "loss/reg": 0.0,
+      "step": 4420
+    },
+    {
+      "epoch": 0.02914473684210526,
+      "grad_norm": 6.375,
+      "grad_norm_var": 1.0507120768229166,
+      "learning_rate": 0.0001,
+      "loss": 3.6517,
+      "loss/crossentropy": 2.547470712661743,
+      "loss/hidden": 3.059375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.298332154750824,
+      "loss/reg": 0.0,
+      "step": 4430
+    },
+    {
+      "epoch": 0.029210526315789475,
+      "grad_norm": 3.796875,
+      "grad_norm_var": 1.0804972330729166,
+      "learning_rate": 0.0001,
+      "loss": 3.7389,
+      "loss/crossentropy": 2.7002538442611694,
+      "loss/hidden": 3.1921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3686490625143051,
+      "loss/reg": 0.0,
+      "step": 4440
+    },
+    {
+      "epoch": 0.029276315789473685,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.432177734375,
+      "learning_rate": 0.0001,
+      "loss": 3.5541,
+      "loss/crossentropy": 2.421727478504181,
+      "loss/hidden": 3.18125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30692713260650634,
+      "loss/reg": 0.0,
+      "step": 4450
+    },
+    {
+      "epoch": 0.029342105263157895,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.09047749837239584,
+      "learning_rate": 0.0001,
+      "loss": 3.6132,
+      "loss/crossentropy": 2.659491038322449,
+      "loss/hidden": 3.2140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3632184773683548,
+      "loss/reg": 0.0,
+      "step": 4460
+    },
+    {
+      "epoch": 0.029407894736842105,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.06815999348958333,
+      "learning_rate": 0.0001,
+      "loss": 3.4561,
+      "loss/crossentropy": 2.5192266911268235,
+      "loss/hidden": 3.034375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2842270269989967,
+      "loss/reg": 0.0,
+      "step": 4470
+    },
+    {
+      "epoch": 0.029473684210526315,
+      "grad_norm": 3.109375,
+      "grad_norm_var": 0.1687408447265625,
+      "learning_rate": 0.0001,
+      "loss": 3.5582,
+      "loss/crossentropy": 2.2021409273147583,
+      "loss/hidden": 3.1484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31400761008262634,
+      "loss/reg": 0.0,
+      "step": 4480
+    },
+    {
+      "epoch": 0.029539473684210525,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.5006174723307292,
+      "learning_rate": 0.0001,
+      "loss": 3.5183,
+      "loss/crossentropy": 2.4442033648490904,
+      "loss/hidden": 3.2640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3307736128568649,
+      "loss/reg": 0.0,
+      "step": 4490
+    },
+    {
+      "epoch": 0.029605263157894735,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.40276692708333334,
+      "learning_rate": 0.0001,
+      "loss": 3.4888,
+      "loss/crossentropy": 2.373897171020508,
+      "loss/hidden": 3.2234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31207202970981596,
+      "loss/reg": 0.0,
+      "step": 4500
+    },
+    {
+      "epoch": 0.02967105263157895,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.20685933430989584,
+      "learning_rate": 0.0001,
+      "loss": 3.5971,
+      "loss/crossentropy": 2.325096046924591,
+      "loss/hidden": 3.40625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4359890788793564,
+      "loss/reg": 0.0,
+      "step": 4510
+    },
+    {
+      "epoch": 0.02973684210526316,
+      "grad_norm": 3.078125,
+      "grad_norm_var": 0.2956451416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.5023,
+      "loss/crossentropy": 2.1537609457969666,
+      "loss/hidden": 3.253125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3213866874575615,
+      "loss/reg": 0.0,
+      "step": 4520
+    },
+    {
+      "epoch": 0.02980263157894737,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.09353841145833333,
+      "learning_rate": 0.0001,
+      "loss": 3.4494,
+      "loss/crossentropy": 2.4694852471351623,
+      "loss/hidden": 2.996875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3156100481748581,
+      "loss/reg": 0.0,
+      "step": 4530
+    },
+    {
+      "epoch": 0.02986842105263158,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.17408447265625,
+      "learning_rate": 0.0001,
+      "loss": 3.6124,
+      "loss/crossentropy": 2.438071775436401,
+      "loss/hidden": 3.3203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40948416888713834,
+      "loss/reg": 0.0,
+      "step": 4540
+    },
+    {
+      "epoch": 0.02993421052631579,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.6079010009765625,
+      "learning_rate": 0.0001,
+      "loss": 3.6751,
+      "loss/crossentropy": 2.525905132293701,
+      "loss/hidden": 3.3546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4153590425848961,
+      "loss/reg": 0.0,
+      "step": 4550
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.6879191080729167,
+      "learning_rate": 0.0001,
+      "loss": 3.5335,
+      "loss/crossentropy": 2.421697771549225,
+      "loss/hidden": 3.025,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3177122876048088,
+      "loss/reg": 0.0,
+      "step": 4560
+    },
+    {
+      "epoch": 0.03006578947368421,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.5993123372395833,
+      "learning_rate": 0.0001,
+      "loss": 3.5742,
+      "loss/crossentropy": 2.3068729996681214,
+      "loss/hidden": 3.05,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29430699050426484,
+      "loss/reg": 0.0,
+      "step": 4570
+    },
+    {
+      "epoch": 0.030131578947368422,
+      "grad_norm": 4.21875,
+      "grad_norm_var": 0.5433339436848958,
+      "learning_rate": 0.0001,
+      "loss": 3.6381,
+      "loss/crossentropy": 2.3981791496276856,
+      "loss/hidden": 3.2359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3355312556028366,
+      "loss/reg": 0.0,
+      "step": 4580
+    },
+    {
+      "epoch": 0.030197368421052632,
+      "grad_norm": 3.25,
+      "grad_norm_var": 0.9806630452473958,
+      "learning_rate": 0.0001,
+      "loss": 3.6522,
+      "loss/crossentropy": 2.309436595439911,
+      "loss/hidden": 3.4890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.350050950050354,
+      "loss/reg": 0.0,
+      "step": 4590
+    },
+    {
+      "epoch": 0.030263157894736843,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.9801910400390625,
+      "learning_rate": 0.0001,
+      "loss": 3.5648,
+      "loss/crossentropy": 2.561086916923523,
+      "loss/hidden": 3.1203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3277510732412338,
+      "loss/reg": 0.0,
+      "step": 4600
+    },
+    {
+      "epoch": 0.030328947368421053,
+      "grad_norm": 3.015625,
+      "grad_norm_var": 0.20364481608072918,
+      "learning_rate": 0.0001,
+      "loss": 3.5518,
+      "loss/crossentropy": 2.6774720311164857,
+      "loss/hidden": 3.0453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31699982583522796,
+      "loss/reg": 0.0,
+      "step": 4610
+    },
+    {
+      "epoch": 0.030394736842105263,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.9744425455729167,
+      "learning_rate": 0.0001,
+      "loss": 3.5924,
+      "loss/crossentropy": 2.623241698741913,
+      "loss/hidden": 3.228125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38138356506824495,
+      "loss/reg": 0.0,
+      "step": 4620
+    },
+    {
+      "epoch": 0.030460526315789473,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.9736073811848959,
+      "learning_rate": 0.0001,
+      "loss": 3.4682,
+      "loss/crossentropy": 2.4855759739875793,
+      "loss/hidden": 3.2671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3377710849046707,
+      "loss/reg": 0.0,
+      "step": 4630
+    },
+    {
+      "epoch": 0.030526315789473683,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.6709706624348958,
+      "learning_rate": 0.0001,
+      "loss": 3.4714,
+      "loss/crossentropy": 2.476264202594757,
+      "loss/hidden": 3.0859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3642516300082207,
+      "loss/reg": 0.0,
+      "step": 4640
+    },
+    {
+      "epoch": 0.030592105263157896,
+      "grad_norm": 3.0,
+      "grad_norm_var": 0.6482899983723959,
+      "learning_rate": 0.0001,
+      "loss": 3.4865,
+      "loss/crossentropy": 2.3190789937973024,
+      "loss/hidden": 3.059375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32634713053703307,
+      "loss/reg": 0.0,
+      "step": 4650
+    },
+    {
+      "epoch": 0.030657894736842106,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.12795817057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.5003,
+      "loss/crossentropy": 2.4210057139396666,
+      "loss/hidden": 3.078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2987945884466171,
+      "loss/reg": 0.0,
+      "step": 4660
+    },
+    {
+      "epoch": 0.030723684210526316,
+      "grad_norm": 3.0625,
+      "grad_norm_var": 0.08050130208333334,
+      "learning_rate": 0.0001,
+      "loss": 3.4858,
+      "loss/crossentropy": 2.041215348243713,
+      "loss/hidden": 3.1953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2908504828810692,
+      "loss/reg": 0.0,
+      "step": 4670
+    },
+    {
+      "epoch": 0.030789473684210526,
+      "grad_norm": 3.0625,
+      "grad_norm_var": 0.4143300374348958,
+      "learning_rate": 0.0001,
+      "loss": 3.6937,
+      "loss/crossentropy": 2.4882567286491395,
+      "loss/hidden": 3.3125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3625297635793686,
+      "loss/reg": 0.0,
+      "step": 4680
+    },
+    {
+      "epoch": 0.030855263157894736,
+      "grad_norm": 3.328125,
+      "grad_norm_var": 0.15123291015625,
+      "learning_rate": 0.0001,
+      "loss": 3.5943,
+      "loss/crossentropy": 2.3254539489746096,
+      "loss/hidden": 3.2,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.308133128285408,
+      "loss/reg": 0.0,
+      "step": 4690
+    },
+    {
+      "epoch": 0.030921052631578946,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.19954325358072916,
+      "learning_rate": 0.0001,
+      "loss": 3.5843,
+      "loss/crossentropy": 1.918275660276413,
+      "loss/hidden": 3.4046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36902148872613905,
+      "loss/reg": 0.0,
+      "step": 4700
+    },
+    {
+      "epoch": 0.030986842105263156,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.04302978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.5845,
+      "loss/crossentropy": 2.573339414596558,
+      "loss/hidden": 3.1625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36042743623256684,
+      "loss/reg": 0.0,
+      "step": 4710
+    },
+    {
+      "epoch": 0.03105263157894737,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.0475982666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.5453,
+      "loss/crossentropy": 2.313191366195679,
+      "loss/hidden": 3.2,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3170273721218109,
+      "loss/reg": 0.0,
+      "step": 4720
+    },
+    {
+      "epoch": 0.03111842105263158,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.0694244384765625,
+      "learning_rate": 0.0001,
+      "loss": 3.4641,
+      "loss/crossentropy": 2.5151350021362306,
+      "loss/hidden": 3.071875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3049825429916382,
+      "loss/reg": 0.0,
+      "step": 4730
+    },
+    {
+      "epoch": 0.03118421052631579,
+      "grad_norm": 4.25,
+      "grad_norm_var": 1.9115549723307292,
+      "learning_rate": 0.0001,
+      "loss": 3.6088,
+      "loss/crossentropy": 2.278876805305481,
+      "loss/hidden": 3.3796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38837724179029465,
+      "loss/reg": 0.0,
+      "step": 4740
+    },
+    {
+      "epoch": 0.03125,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 2.200390625,
+      "learning_rate": 0.0001,
+      "loss": 3.5348,
+      "loss/crossentropy": 2.2090991735458374,
+      "loss/hidden": 3.44375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3878710061311722,
+      "loss/reg": 0.0,
+      "step": 4750
+    },
+    {
+      "epoch": 0.031315789473684214,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.6573404947916667,
+      "learning_rate": 0.0001,
+      "loss": 3.5378,
+      "loss/crossentropy": 2.2805041670799255,
+      "loss/hidden": 3.3015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31536445766687393,
+      "loss/reg": 0.0,
+      "step": 4760
+    },
+    {
+      "epoch": 0.03138157894736842,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.38818257649739585,
+      "learning_rate": 0.0001,
+      "loss": 3.6082,
+      "loss/crossentropy": 2.6178433656692506,
+      "loss/hidden": 3.2828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3645938545465469,
+      "loss/reg": 0.0,
+      "step": 4770
+    },
+    {
+      "epoch": 0.031447368421052634,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.07810770670572917,
+      "learning_rate": 0.0001,
+      "loss": 3.4424,
+      "loss/crossentropy": 2.5498223304748535,
+      "loss/hidden": 3.2625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3333883464336395,
+      "loss/reg": 0.0,
+      "step": 4780
+    },
+    {
+      "epoch": 0.03151315789473684,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.11669820149739583,
+      "learning_rate": 0.0001,
+      "loss": 3.6145,
+      "loss/crossentropy": 2.639970850944519,
+      "loss/hidden": 3.278125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37047617733478544,
+      "loss/reg": 0.0,
+      "step": 4790
+    },
+    {
+      "epoch": 0.031578947368421054,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.0907867431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.5869,
+      "loss/crossentropy": 2.458595323562622,
+      "loss/hidden": 3.60625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4149588346481323,
+      "loss/reg": 0.0,
+      "step": 4800
+    },
+    {
+      "epoch": 0.03164473684210526,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.0759429931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.4775,
+      "loss/crossentropy": 2.01421400308609,
+      "loss/hidden": 3.2671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29176320880651474,
+      "loss/reg": 0.0,
+      "step": 4810
+    },
+    {
+      "epoch": 0.031710526315789474,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.1147857666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.5268,
+      "loss/crossentropy": 2.5456383228302,
+      "loss/hidden": 3.3484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3915561467409134,
+      "loss/reg": 0.0,
+      "step": 4820
+    },
+    {
+      "epoch": 0.03177631578947369,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.0890625,
+      "learning_rate": 0.0001,
+      "loss": 3.5371,
+      "loss/crossentropy": 2.1736844003200533,
+      "loss/hidden": 3.0078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28229199200868604,
+      "loss/reg": 0.0,
+      "step": 4830
+    },
+    {
+      "epoch": 0.031842105263157894,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.08162434895833333,
+      "learning_rate": 0.0001,
+      "loss": 3.4898,
+      "loss/crossentropy": 2.397980511188507,
+      "loss/hidden": 3.06875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28644354790449145,
+      "loss/reg": 0.0,
+      "step": 4840
+    },
+    {
+      "epoch": 0.03190789473684211,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 1.33443603515625,
+      "learning_rate": 0.0001,
+      "loss": 3.5788,
+      "loss/crossentropy": 2.434200632572174,
+      "loss/hidden": 3.103125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37485773116350174,
+      "loss/reg": 0.0,
+      "step": 4850
+    },
+    {
+      "epoch": 0.031973684210526314,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 1.3780181884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.5008,
+      "loss/crossentropy": 2.642025816440582,
+      "loss/hidden": 3.1375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34916335344314575,
+      "loss/reg": 0.0,
+      "step": 4860
+    },
+    {
+      "epoch": 0.03203947368421053,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.04523111979166667,
+      "learning_rate": 0.0001,
+      "loss": 3.4784,
+      "loss/crossentropy": 2.2365106463432314,
+      "loss/hidden": 3.4328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.45036998838186265,
+      "loss/reg": 0.0,
+      "step": 4870
+    },
+    {
+      "epoch": 0.032105263157894734,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.04436442057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.5047,
+      "loss/crossentropy": 2.250430929660797,
+      "loss/hidden": 3.2859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3373000741004944,
+      "loss/reg": 0.0,
+      "step": 4880
+    },
+    {
+      "epoch": 0.03217105263157895,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.07979227701822916,
+      "learning_rate": 0.0001,
+      "loss": 3.4476,
+      "loss/crossentropy": 2.688676381111145,
+      "loss/hidden": 3.184375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32961316406726837,
+      "loss/reg": 0.0,
+      "step": 4890
+    },
+    {
+      "epoch": 0.03223684210526316,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.0539215087890625,
+      "learning_rate": 0.0001,
+      "loss": 3.5264,
+      "loss/crossentropy": 2.5281107783317567,
+      "loss/hidden": 3.2125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36207843720912936,
+      "loss/reg": 0.0,
+      "step": 4900
+    },
+    {
+      "epoch": 0.03230263157894737,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.11297200520833334,
+      "learning_rate": 0.0001,
+      "loss": 3.4276,
+      "loss/crossentropy": 2.3438509345054626,
+      "loss/hidden": 3.2671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3646134212613106,
+      "loss/reg": 0.0,
+      "step": 4910
+    },
+    {
+      "epoch": 0.03236842105263158,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.06402587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.479,
+      "loss/crossentropy": 2.331527066230774,
+      "loss/hidden": 3.48125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.407479552924633,
+      "loss/reg": 0.0,
+      "step": 4920
+    },
+    {
+      "epoch": 0.03243421052631579,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.21721598307291667,
+      "learning_rate": 0.0001,
+      "loss": 3.6251,
+      "loss/crossentropy": 2.3732258677482605,
+      "loss/hidden": 3.1875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32182002663612364,
+      "loss/reg": 0.0,
+      "step": 4930
+    },
+    {
+      "epoch": 0.0325,
+      "grad_norm": 4.59375,
+      "grad_norm_var": 1.3630167643229167,
+      "learning_rate": 0.0001,
+      "loss": 3.6469,
+      "loss/crossentropy": 1.9915230482816697,
+      "loss/hidden": 3.3953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35692891776561736,
+      "loss/reg": 0.0,
+      "step": 4940
+    },
+    {
+      "epoch": 0.03256578947368421,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.2885894775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.5077,
+      "loss/crossentropy": 1.9868581891059875,
+      "loss/hidden": 3.215625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30731415897607806,
+      "loss/reg": 0.0,
+      "step": 4950
+    },
+    {
+      "epoch": 0.03263157894736842,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.21422119140625,
+      "learning_rate": 0.0001,
+      "loss": 3.5426,
+      "loss/crossentropy": 2.3579143285751343,
+      "loss/hidden": 3.3421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3354805111885071,
+      "loss/reg": 0.0,
+      "step": 4960
+    },
+    {
+      "epoch": 0.032697368421052635,
+      "grad_norm": 2.625,
+      "grad_norm_var": 4.145735677083334,
+      "learning_rate": 0.0001,
+      "loss": 3.5276,
+      "loss/crossentropy": 2.2787723779678344,
+      "loss/hidden": 2.9453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2516452088952065,
+      "loss/reg": 0.0,
+      "step": 4970
+    },
+    {
+      "epoch": 0.03276315789473684,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.09233296712239583,
+      "learning_rate": 0.0001,
+      "loss": 3.4912,
+      "loss/crossentropy": 2.2899803042411806,
+      "loss/hidden": 3.175,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3140288829803467,
+      "loss/reg": 0.0,
+      "step": 4980
+    },
+    {
+      "epoch": 0.032828947368421055,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.07078450520833333,
+      "learning_rate": 0.0001,
+      "loss": 3.4308,
+      "loss/crossentropy": 2.4203084468841554,
+      "loss/hidden": 3.009375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2926447048783302,
+      "loss/reg": 0.0,
+      "step": 4990
+    },
+    {
+      "epoch": 0.03289473684210526,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.1664215087890625,
+      "learning_rate": 0.0001,
+      "loss": 3.5233,
+      "loss/crossentropy": 2.4435134291648866,
+      "loss/hidden": 3.3640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35478622317314146,
+      "loss/reg": 0.0,
+      "step": 5000
+    },
+    {
+      "epoch": 0.032960526315789475,
+      "grad_norm": 3.296875,
+      "grad_norm_var": 0.17375386555989583,
+      "learning_rate": 0.0001,
+      "loss": 3.529,
+      "loss/crossentropy": 2.3886643409729005,
+      "loss/hidden": 3.196875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36350963413715365,
+      "loss/reg": 0.0,
+      "step": 5010
+    },
+    {
+      "epoch": 0.03302631578947368,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.13043619791666666,
+      "learning_rate": 0.0001,
+      "loss": 3.5608,
+      "loss/crossentropy": 2.5570758461952208,
+      "loss/hidden": 3.4828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.343785697221756,
+      "loss/reg": 0.0,
+      "step": 5020
+    },
+    {
+      "epoch": 0.033092105263157895,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.19719645182291667,
+      "learning_rate": 0.0001,
+      "loss": 3.5903,
+      "loss/crossentropy": 2.3763694763183594,
+      "loss/hidden": 3.25625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32882467210292815,
+      "loss/reg": 0.0,
+      "step": 5030
+    },
+    {
+      "epoch": 0.03315789473684211,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.21155497233072917,
+      "learning_rate": 0.0001,
+      "loss": 3.5454,
+      "loss/crossentropy": 2.5775513648986816,
+      "loss/hidden": 3.0375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3086023017764091,
+      "loss/reg": 0.0,
+      "step": 5040
+    },
+    {
+      "epoch": 0.033223684210526315,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.08268229166666667,
+      "learning_rate": 0.0001,
+      "loss": 3.5005,
+      "loss/crossentropy": 2.257374918460846,
+      "loss/hidden": 3.1609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3117083102464676,
+      "loss/reg": 0.0,
+      "step": 5050
+    },
+    {
+      "epoch": 0.03328947368421053,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.21896158854166667,
+      "learning_rate": 0.0001,
+      "loss": 3.4097,
+      "loss/crossentropy": 2.437604343891144,
+      "loss/hidden": 3.1953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33744728565216064,
+      "loss/reg": 0.0,
+      "step": 5060
+    },
+    {
+      "epoch": 0.033355263157894735,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.19402669270833334,
+      "learning_rate": 0.0001,
+      "loss": 3.5413,
+      "loss/crossentropy": 2.1056251645088198,
+      "loss/hidden": 3.3140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35452440977096555,
+      "loss/reg": 0.0,
+      "step": 5070
+    },
+    {
+      "epoch": 0.03342105263157895,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 0.041829427083333336,
+      "learning_rate": 0.0001,
+      "loss": 3.5305,
+      "loss/crossentropy": 2.5163299083709716,
+      "loss/hidden": 3.296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35606471002101897,
+      "loss/reg": 0.0,
+      "step": 5080
+    },
+    {
+      "epoch": 0.033486842105263155,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.08046773274739584,
+      "learning_rate": 0.0001,
+      "loss": 3.4405,
+      "loss/crossentropy": 2.37408185005188,
+      "loss/hidden": 2.9828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3234561800956726,
+      "loss/reg": 0.0,
+      "step": 5090
+    },
+    {
+      "epoch": 0.03355263157894737,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.17653706868489583,
+      "learning_rate": 0.0001,
+      "loss": 3.4303,
+      "loss/crossentropy": 2.3730108022689818,
+      "loss/hidden": 3.1359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2919617787003517,
+      "loss/reg": 0.0,
+      "step": 5100
+    },
+    {
+      "epoch": 0.03361842105263158,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.47526753743489586,
+      "learning_rate": 0.0001,
+      "loss": 3.5312,
+      "loss/crossentropy": 2.4720141887664795,
+      "loss/hidden": 3.246875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36764703392982484,
+      "loss/reg": 0.0,
+      "step": 5110
+    },
+    {
+      "epoch": 0.03368421052631579,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.4923787434895833,
+      "learning_rate": 0.0001,
+      "loss": 3.5491,
+      "loss/crossentropy": 2.449038088321686,
+      "loss/hidden": 3.3171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4042062431573868,
+      "loss/reg": 0.0,
+      "step": 5120
+    },
+    {
+      "epoch": 0.03375,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.1626861572265625,
+      "learning_rate": 0.0001,
+      "loss": 3.5298,
+      "loss/crossentropy": 2.5271955728530884,
+      "loss/hidden": 3.2515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3406914800405502,
+      "loss/reg": 0.0,
+      "step": 5130
+    },
+    {
+      "epoch": 0.03381578947368421,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.037018839518229166,
+      "learning_rate": 0.0001,
+      "loss": 3.549,
+      "loss/crossentropy": 2.6082807898521425,
+      "loss/hidden": 3.321875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33228414356708524,
+      "loss/reg": 0.0,
+      "step": 5140
+    },
+    {
+      "epoch": 0.03388157894736842,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.08620503743489584,
+      "learning_rate": 0.0001,
+      "loss": 3.505,
+      "loss/crossentropy": 2.0589061468839644,
+      "loss/hidden": 3.4890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30756633579730985,
+      "loss/reg": 0.0,
+      "step": 5150
+    },
+    {
+      "epoch": 0.03394736842105263,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.3478342692057292,
+      "learning_rate": 0.0001,
+      "loss": 3.5598,
+      "loss/crossentropy": 2.0876080930233,
+      "loss/hidden": 3.39375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3331515982747078,
+      "loss/reg": 0.0,
+      "step": 5160
+    },
+    {
+      "epoch": 0.03401315789473684,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.3102773030598958,
+      "learning_rate": 0.0001,
+      "loss": 3.5298,
+      "loss/crossentropy": 2.123845911026001,
+      "loss/hidden": 3.2953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27716329991817473,
+      "loss/reg": 0.0,
+      "step": 5170
+    },
+    {
+      "epoch": 0.034078947368421056,
+      "grad_norm": 3.484375,
+      "grad_norm_var": 0.3398590087890625,
+      "learning_rate": 0.0001,
+      "loss": 3.546,
+      "loss/crossentropy": 2.5750380873680117,
+      "loss/hidden": 3.0671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3079290196299553,
+      "loss/reg": 0.0,
+      "step": 5180
+    },
+    {
+      "epoch": 0.03414473684210526,
+      "grad_norm": 3.015625,
+      "grad_norm_var": 0.3519195556640625,
+      "learning_rate": 0.0001,
+      "loss": 3.4686,
+      "loss/crossentropy": 2.1761133074760437,
+      "loss/hidden": 3.0421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28422126173973083,
+      "loss/reg": 0.0,
+      "step": 5190
+    },
+    {
+      "epoch": 0.034210526315789476,
+      "grad_norm": 3.21875,
+      "grad_norm_var": 0.09970296223958333,
+      "learning_rate": 0.0001,
+      "loss": 3.5823,
+      "loss/crossentropy": 2.224585694074631,
+      "loss/hidden": 3.2109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32807315289974215,
+      "loss/reg": 0.0,
+      "step": 5200
+    },
+    {
+      "epoch": 0.03427631578947368,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.34479878743489584,
+      "learning_rate": 0.0001,
+      "loss": 3.5417,
+      "loss/crossentropy": 2.217251694202423,
+      "loss/hidden": 3.159375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2984598934650421,
+      "loss/reg": 0.0,
+      "step": 5210
+    },
+    {
+      "epoch": 0.034342105263157896,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.28804931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.4788,
+      "loss/crossentropy": 2.344852977991104,
+      "loss/hidden": 3.1265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2878506749868393,
+      "loss/reg": 0.0,
+      "step": 5220
+    },
+    {
+      "epoch": 0.0344078947368421,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.17006734212239583,
+      "learning_rate": 0.0001,
+      "loss": 3.4649,
+      "loss/crossentropy": 2.5100401520729063,
+      "loss/hidden": 3.128125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3491516515612602,
+      "loss/reg": 0.0,
+      "step": 5230
+    },
+    {
+      "epoch": 0.034473684210526316,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 1.1076649983723958,
+      "learning_rate": 0.0001,
+      "loss": 3.503,
+      "loss/crossentropy": 2.565778684616089,
+      "loss/hidden": 3.146875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35944747030735014,
+      "loss/reg": 0.0,
+      "step": 5240
+    },
+    {
+      "epoch": 0.03453947368421053,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 1.1434315999348958,
+      "learning_rate": 0.0001,
+      "loss": 3.4668,
+      "loss/crossentropy": 2.5031490683555604,
+      "loss/hidden": 2.99375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2864773109555244,
+      "loss/reg": 0.0,
+      "step": 5250
+    },
+    {
+      "epoch": 0.034605263157894736,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.5837849934895833,
+      "learning_rate": 0.0001,
+      "loss": 3.4822,
+      "loss/crossentropy": 2.3741963386535643,
+      "loss/hidden": 3.2078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3231631726026535,
+      "loss/reg": 0.0,
+      "step": 5260
+    },
+    {
+      "epoch": 0.03467105263157895,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.62431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.4414,
+      "loss/crossentropy": 2.3789267897605897,
+      "loss/hidden": 3.1921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3201348423957825,
+      "loss/reg": 0.0,
+      "step": 5270
+    },
+    {
+      "epoch": 0.034736842105263156,
+      "grad_norm": 5.25,
+      "grad_norm_var": 0.5048573811848959,
+      "learning_rate": 0.0001,
+      "loss": 3.412,
+      "loss/crossentropy": 2.47695529460907,
+      "loss/hidden": 3.171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31060084253549575,
+      "loss/reg": 0.0,
+      "step": 5280
+    },
+    {
+      "epoch": 0.03480263157894737,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.54195556640625,
+      "learning_rate": 0.0001,
+      "loss": 3.5371,
+      "loss/crossentropy": 2.4316977143287657,
+      "loss/hidden": 3.0703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30330550074577334,
+      "loss/reg": 0.0,
+      "step": 5290
+    },
+    {
+      "epoch": 0.034868421052631576,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.1771484375,
+      "learning_rate": 0.0001,
+      "loss": 3.534,
+      "loss/crossentropy": 2.058604693412781,
+      "loss/hidden": 3.16875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2998970851302147,
+      "loss/reg": 0.0,
+      "step": 5300
+    },
+    {
+      "epoch": 0.03493421052631579,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.21590169270833334,
+      "learning_rate": 0.0001,
+      "loss": 3.5029,
+      "loss/crossentropy": 2.1623964309692383,
+      "loss/hidden": 3.1453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2633717767894268,
+      "loss/reg": 0.0,
+      "step": 5310
+    },
+    {
+      "epoch": 0.035,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.12704671223958333,
+      "learning_rate": 0.0001,
+      "loss": 3.4862,
+      "loss/crossentropy": 2.5717769265174866,
+      "loss/hidden": 3.1671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30162925869226453,
+      "loss/reg": 0.0,
+      "step": 5320
+    },
+    {
+      "epoch": 0.03506578947368421,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 1.4618398030598958,
+      "learning_rate": 0.0001,
+      "loss": 3.5429,
+      "loss/crossentropy": 2.462851893901825,
+      "loss/hidden": 3.328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3766929477453232,
+      "loss/reg": 0.0,
+      "step": 5330
+    },
+    {
+      "epoch": 0.03513157894736842,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 1.7596181233723958,
+      "learning_rate": 0.0001,
+      "loss": 3.5089,
+      "loss/crossentropy": 2.44319885969162,
+      "loss/hidden": 3.034375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30458342134952543,
+      "loss/reg": 0.0,
+      "step": 5340
+    },
+    {
+      "epoch": 0.03519736842105263,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.4787668863932292,
+      "learning_rate": 0.0001,
+      "loss": 3.6272,
+      "loss/crossentropy": 2.594151020050049,
+      "loss/hidden": 3.2296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3736713409423828,
+      "loss/reg": 0.0,
+      "step": 5350
+    },
+    {
+      "epoch": 0.035263157894736843,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.4522939046223958,
+      "learning_rate": 0.0001,
+      "loss": 3.5325,
+      "loss/crossentropy": 2.0771877110004424,
+      "loss/hidden": 3.4046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3338633939623833,
+      "loss/reg": 0.0,
+      "step": 5360
+    },
+    {
+      "epoch": 0.03532894736842105,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.0387847900390625,
+      "learning_rate": 0.0001,
+      "loss": 3.4757,
+      "loss/crossentropy": 2.5547770977020265,
+      "loss/hidden": 3.25,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36584808975458144,
+      "loss/reg": 0.0,
+      "step": 5370
+    },
+    {
+      "epoch": 0.035394736842105264,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 11.889094034830729,
+      "learning_rate": 0.0001,
+      "loss": 3.6332,
+      "loss/crossentropy": 2.2124004304409026,
+      "loss/hidden": 3.1265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29226877391338346,
+      "loss/reg": 0.0,
+      "step": 5380
+    },
+    {
+      "epoch": 0.03546052631578948,
+      "grad_norm": 3.15625,
+      "grad_norm_var": 6.872362263997396,
+      "learning_rate": 0.0001,
+      "loss": 3.5496,
+      "loss/crossentropy": 2.2622018218040467,
+      "loss/hidden": 3.259375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40298803299665453,
+      "loss/reg": 0.0,
+      "step": 5390
+    },
+    {
+      "epoch": 0.035526315789473684,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.0641998291015625,
+      "learning_rate": 0.0001,
+      "loss": 3.4621,
+      "loss/crossentropy": 2.374979627132416,
+      "loss/hidden": 3.2109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37631402611732484,
+      "loss/reg": 0.0,
+      "step": 5400
+    },
+    {
+      "epoch": 0.0355921052631579,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.028059895833333334,
+      "learning_rate": 0.0001,
+      "loss": 3.4667,
+      "loss/crossentropy": 2.4809056520462036,
+      "loss/hidden": 3.1078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32154888212680816,
+      "loss/reg": 0.0,
+      "step": 5410
+    },
+    {
+      "epoch": 0.035657894736842104,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.19182840983072916,
+      "learning_rate": 0.0001,
+      "loss": 3.528,
+      "loss/crossentropy": 2.3937729835510253,
+      "loss/hidden": 3.1234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29925636053085325,
+      "loss/reg": 0.0,
+      "step": 5420
+    },
+    {
+      "epoch": 0.03572368421052632,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.04440104166666667,
+      "learning_rate": 0.0001,
+      "loss": 3.4433,
+      "loss/crossentropy": 2.604015350341797,
+      "loss/hidden": 2.978125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28630980402231215,
+      "loss/reg": 0.0,
+      "step": 5430
+    },
+    {
+      "epoch": 0.035789473684210524,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.057454427083333336,
+      "learning_rate": 0.0001,
+      "loss": 3.3938,
+      "loss/crossentropy": 2.3647801518440246,
+      "loss/hidden": 3.1421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2772139713168144,
+      "loss/reg": 0.0,
+      "step": 5440
+    },
+    {
+      "epoch": 0.03585526315789474,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.06382548014322917,
+      "learning_rate": 0.0001,
+      "loss": 3.4874,
+      "loss/crossentropy": 2.4227387428283693,
+      "loss/hidden": 3.3890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3498344630002975,
+      "loss/reg": 0.0,
+      "step": 5450
+    },
+    {
+      "epoch": 0.03592105263157895,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.5879221598307292,
+      "learning_rate": 0.0001,
+      "loss": 3.5541,
+      "loss/crossentropy": 2.024171155691147,
+      "loss/hidden": 3.003125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25232082083821294,
+      "loss/reg": 0.0,
+      "step": 5460
+    },
+    {
+      "epoch": 0.03598684210526316,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 1.2988596598307292,
+      "learning_rate": 0.0001,
+      "loss": 3.4549,
+      "loss/crossentropy": 2.5548394203186033,
+      "loss/hidden": 3.0234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29853117763996123,
+      "loss/reg": 0.0,
+      "step": 5470
+    },
+    {
+      "epoch": 0.03605263157894737,
+      "grad_norm": 3.234375,
+      "grad_norm_var": 1.458153279622396,
+      "learning_rate": 0.0001,
+      "loss": 3.5047,
+      "loss/crossentropy": 2.520615005493164,
+      "loss/hidden": 3.075,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31394066512584684,
+      "loss/reg": 0.0,
+      "step": 5480
+    },
+    {
+      "epoch": 0.03611842105263158,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.6611328125,
+      "learning_rate": 0.0001,
+      "loss": 3.4932,
+      "loss/crossentropy": 2.447161090373993,
+      "loss/hidden": 3.090625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3091880366206169,
+      "loss/reg": 0.0,
+      "step": 5490
+    },
+    {
+      "epoch": 0.03618421052631579,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.05611572265625,
+      "learning_rate": 0.0001,
+      "loss": 3.4718,
+      "loss/crossentropy": 2.354436981678009,
+      "loss/hidden": 3.0828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30545540153980255,
+      "loss/reg": 0.0,
+      "step": 5500
+    },
+    {
+      "epoch": 0.03625,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.24016011555989583,
+      "learning_rate": 0.0001,
+      "loss": 3.5554,
+      "loss/crossentropy": 2.350696861743927,
+      "loss/hidden": 3.1578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27360412031412124,
+      "loss/reg": 0.0,
+      "step": 5510
+    },
+    {
+      "epoch": 0.03631578947368421,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.6210245768229167,
+      "learning_rate": 0.0001,
+      "loss": 3.513,
+      "loss/crossentropy": 2.368817460536957,
+      "loss/hidden": 3.1796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31491883993148806,
+      "loss/reg": 0.0,
+      "step": 5520
+    },
+    {
+      "epoch": 0.036381578947368425,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.16715087890625,
+      "learning_rate": 0.0001,
+      "loss": 3.4444,
+      "loss/crossentropy": 2.3894132494926454,
+      "loss/hidden": 3.075,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2710603341460228,
+      "loss/reg": 0.0,
+      "step": 5530
+    },
+    {
+      "epoch": 0.03644736842105263,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 2.993724568684896,
+      "learning_rate": 0.0001,
+      "loss": 3.5105,
+      "loss/crossentropy": 2.4798691868782043,
+      "loss/hidden": 3.25,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.422188438475132,
+      "loss/reg": 0.0,
+      "step": 5540
+    },
+    {
+      "epoch": 0.036513157894736845,
+      "grad_norm": 3.6875,
+      "grad_norm_var": 2.864090983072917,
+      "learning_rate": 0.0001,
+      "loss": 3.6266,
+      "loss/crossentropy": 2.499036192893982,
+      "loss/hidden": 3.21875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3576551049947739,
+      "loss/reg": 0.0,
+      "step": 5550
+    },
+    {
+      "epoch": 0.03657894736842105,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.2598052978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.5659,
+      "loss/crossentropy": 2.4270546317100523,
+      "loss/hidden": 3.23125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4616221562027931,
+      "loss/reg": 0.0,
+      "step": 5560
+    },
+    {
+      "epoch": 0.036644736842105265,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.39485270182291665,
+      "learning_rate": 0.0001,
+      "loss": 3.5978,
+      "loss/crossentropy": 2.427480709552765,
+      "loss/hidden": 3.7875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38075721710920335,
+      "loss/reg": 0.0,
+      "step": 5570
+    },
+    {
+      "epoch": 0.03671052631578947,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 1.5193318684895833,
+      "learning_rate": 0.0001,
+      "loss": 3.5541,
+      "loss/crossentropy": 2.2717662811279298,
+      "loss/hidden": 3.2359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2794697627425194,
+      "loss/reg": 0.0,
+      "step": 5580
+    },
+    {
+      "epoch": 0.036776315789473685,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 1.5052734375,
+      "learning_rate": 0.0001,
+      "loss": 3.5513,
+      "loss/crossentropy": 2.549258255958557,
+      "loss/hidden": 3.1265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3551526039838791,
+      "loss/reg": 0.0,
+      "step": 5590
+    },
+    {
+      "epoch": 0.03684210526315789,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.12431538899739583,
+      "learning_rate": 0.0001,
+      "loss": 3.5166,
+      "loss/crossentropy": 2.42179411649704,
+      "loss/hidden": 3.265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.349351167678833,
+      "loss/reg": 0.0,
+      "step": 5600
+    },
+    {
+      "epoch": 0.036907894736842105,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.09789937337239583,
+      "learning_rate": 0.0001,
+      "loss": 3.4519,
+      "loss/crossentropy": 2.275598430633545,
+      "loss/hidden": 3.2015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32426146864891053,
+      "loss/reg": 0.0,
+      "step": 5610
+    },
+    {
+      "epoch": 0.03697368421052632,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.10132548014322916,
+      "learning_rate": 0.0001,
+      "loss": 3.4632,
+      "loss/crossentropy": 2.4317312955856325,
+      "loss/hidden": 3.3171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3550658613443375,
+      "loss/reg": 0.0,
+      "step": 5620
+    },
+    {
+      "epoch": 0.037039473684210525,
+      "grad_norm": 25.0,
+      "grad_norm_var": 167.39566650390626,
+      "learning_rate": 0.0001,
+      "loss": 3.6136,
+      "loss/crossentropy": 2.5963298320770263,
+      "loss/hidden": 3.1546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34572866559028625,
+      "loss/reg": 0.0,
+      "step": 5630
+    },
+    {
+      "epoch": 0.03710526315789474,
+      "grad_norm": 3.0,
+      "grad_norm_var": 167.50836486816405,
+      "learning_rate": 0.0001,
+      "loss": 3.5122,
+      "loss/crossentropy": 2.370168614387512,
+      "loss/hidden": 3.090625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.316168874502182,
+      "loss/reg": 0.0,
+      "step": 5640
+    },
+    {
+      "epoch": 0.037171052631578945,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.05074462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.4512,
+      "loss/crossentropy": 2.1566815614700316,
+      "loss/hidden": 3.109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30936725735664367,
+      "loss/reg": 0.0,
+      "step": 5650
+    },
+    {
+      "epoch": 0.03723684210526316,
+      "grad_norm": 3.546875,
+      "grad_norm_var": 0.1086822509765625,
+      "learning_rate": 0.0001,
+      "loss": 3.4308,
+      "loss/crossentropy": 2.315059244632721,
+      "loss/hidden": 3.1609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3344813346862793,
+      "loss/reg": 0.0,
+      "step": 5660
+    },
+    {
+      "epoch": 0.037302631578947365,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 0.22506510416666667,
+      "learning_rate": 0.0001,
+      "loss": 3.5705,
+      "loss/crossentropy": 2.063437449932098,
+      "loss/hidden": 3.1421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2783856257796288,
+      "loss/reg": 0.0,
+      "step": 5670
+    },
+    {
+      "epoch": 0.03736842105263158,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.48176981608072916,
+      "learning_rate": 0.0001,
+      "loss": 3.5285,
+      "loss/crossentropy": 2.564136099815369,
+      "loss/hidden": 3.2453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3260859474539757,
+      "loss/reg": 0.0,
+      "step": 5680
+    },
+    {
+      "epoch": 0.03743421052631579,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 1.2526519775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.581,
+      "loss/crossentropy": 2.4384737968444825,
+      "loss/hidden": 2.9515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2700443536043167,
+      "loss/reg": 0.0,
+      "step": 5690
+    },
+    {
+      "epoch": 0.0375,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 1.029296875,
+      "learning_rate": 0.0001,
+      "loss": 3.5309,
+      "loss/crossentropy": 2.656829285621643,
+      "loss/hidden": 3.1328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34218672215938567,
+      "loss/reg": 0.0,
+      "step": 5700
+    },
+    {
+      "epoch": 0.03756578947368421,
+      "grad_norm": 2.375,
+      "grad_norm_var": 5.287238566080729,
+      "learning_rate": 0.0001,
+      "loss": 3.6012,
+      "loss/crossentropy": 2.3537548005580904,
+      "loss/hidden": 3.1875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33707170784473417,
+      "loss/reg": 0.0,
+      "step": 5710
+    },
+    {
+      "epoch": 0.03763157894736842,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.6506795247395833,
+      "learning_rate": 0.0001,
+      "loss": 3.4987,
+      "loss/crossentropy": 2.217307722568512,
+      "loss/hidden": 3.31875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36700052917003634,
+      "loss/reg": 0.0,
+      "step": 5720
+    },
+    {
+      "epoch": 0.03769736842105263,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 235.47136942545572,
+      "learning_rate": 0.0001,
+      "loss": 3.4822,
+      "loss/crossentropy": 2.39666086435318,
+      "loss/hidden": 3.0828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.322082930803299,
+      "loss/reg": 0.0,
+      "step": 5730
+    },
+    {
+      "epoch": 0.03776315789473684,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 48.15474853515625,
+      "learning_rate": 0.0001,
+      "loss": 3.5338,
+      "loss/crossentropy": 2.5715784192085267,
+      "loss/hidden": 3.103125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31636003255844114,
+      "loss/reg": 0.0,
+      "step": 5740
+    },
+    {
+      "epoch": 0.03782894736842105,
+      "grad_norm": 2.375,
+      "grad_norm_var": 48.25117899576823,
+      "learning_rate": 0.0001,
+      "loss": 3.345,
+      "loss/crossentropy": 2.1092816948890687,
+      "loss/hidden": 2.9640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27086046040058137,
+      "loss/reg": 0.0,
+      "step": 5750
+    },
+    {
+      "epoch": 0.037894736842105266,
+      "grad_norm": 3.34375,
+      "grad_norm_var": 0.1339508056640625,
+      "learning_rate": 0.0001,
+      "loss": 3.4478,
+      "loss/crossentropy": 2.20155810713768,
+      "loss/hidden": 3.209375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3250092178583145,
+      "loss/reg": 0.0,
+      "step": 5760
+    },
+    {
+      "epoch": 0.03796052631578947,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.5598592122395833,
+      "learning_rate": 0.0001,
+      "loss": 3.4268,
+      "loss/crossentropy": 2.2270141005516053,
+      "loss/hidden": 2.959375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2702811732888222,
+      "loss/reg": 0.0,
+      "step": 5770
+    },
+    {
+      "epoch": 0.038026315789473686,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.1083892822265625,
+      "learning_rate": 0.0001,
+      "loss": 3.4578,
+      "loss/crossentropy": 2.4509124517440797,
+      "loss/hidden": 3.1109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3049029678106308,
+      "loss/reg": 0.0,
+      "step": 5780
+    },
+    {
+      "epoch": 0.03809210526315789,
+      "grad_norm": 3.21875,
+      "grad_norm_var": 0.20041402180989584,
+      "learning_rate": 0.0001,
+      "loss": 3.4516,
+      "loss/crossentropy": 2.1973756074905397,
+      "loss/hidden": 2.946875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2764800027012825,
+      "loss/reg": 0.0,
+      "step": 5790
+    },
+    {
+      "epoch": 0.038157894736842106,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.17437235514322916,
+      "learning_rate": 0.0001,
+      "loss": 3.3789,
+      "loss/crossentropy": 2.4859437584877013,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28125611394643785,
+      "loss/reg": 0.0,
+      "step": 5800
+    },
+    {
+      "epoch": 0.03822368421052631,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 1.9865234375,
+      "learning_rate": 0.0001,
+      "loss": 3.4479,
+      "loss/crossentropy": 2.6306315779685976,
+      "loss/hidden": 3.440625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2930992156267166,
+      "loss/reg": 0.0,
+      "step": 5810
+    },
+    {
+      "epoch": 0.038289473684210526,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 1.902880859375,
+      "learning_rate": 0.0001,
+      "loss": 3.482,
+      "loss/crossentropy": 2.4275772333145142,
+      "loss/hidden": 3.0953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32571674734354017,
+      "loss/reg": 0.0,
+      "step": 5820
+    },
+    {
+      "epoch": 0.03835526315789474,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.040087890625,
+      "learning_rate": 0.0001,
+      "loss": 3.3832,
+      "loss/crossentropy": 2.348308402299881,
+      "loss/hidden": 3.140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.325964193046093,
+      "loss/reg": 0.0,
+      "step": 5830
+    },
+    {
+      "epoch": 0.038421052631578946,
+      "grad_norm": 3.796875,
+      "grad_norm_var": 0.8002115885416666,
+      "learning_rate": 0.0001,
+      "loss": 3.5519,
+      "loss/crossentropy": 2.1252057909965516,
+      "loss/hidden": 3.2984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3046886622905731,
+      "loss/reg": 0.0,
+      "step": 5840
+    },
+    {
+      "epoch": 0.03848684210526316,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.2637685139973958,
+      "learning_rate": 0.0001,
+      "loss": 3.4841,
+      "loss/crossentropy": 2.5330613613128663,
+      "loss/hidden": 3.025,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31516623198986055,
+      "loss/reg": 0.0,
+      "step": 5850
+    },
+    {
+      "epoch": 0.038552631578947366,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.24798075358072916,
+      "learning_rate": 0.0001,
+      "loss": 3.4625,
+      "loss/crossentropy": 2.324964237213135,
+      "loss/hidden": 3.0921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29473926275968554,
+      "loss/reg": 0.0,
+      "step": 5860
+    },
+    {
+      "epoch": 0.03861842105263158,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.12890218098958334,
+      "learning_rate": 0.0001,
+      "loss": 3.3932,
+      "loss/crossentropy": 2.436046540737152,
+      "loss/hidden": 2.9859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28259717375040055,
+      "loss/reg": 0.0,
+      "step": 5870
+    },
+    {
+      "epoch": 0.038684210526315786,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.09658101399739584,
+      "learning_rate": 0.0001,
+      "loss": 3.4335,
+      "loss/crossentropy": 2.3942569494247437,
+      "loss/hidden": 3.040625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29623564779758454,
+      "loss/reg": 0.0,
+      "step": 5880
+    },
+    {
+      "epoch": 0.03875,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.0283355712890625,
+      "learning_rate": 0.0001,
+      "loss": 3.4179,
+      "loss/crossentropy": 2.675841474533081,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2852372720837593,
+      "loss/reg": 0.0,
+      "step": 5890
+    },
+    {
+      "epoch": 0.03881578947368421,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.25607096354166664,
+      "learning_rate": 0.0001,
+      "loss": 3.4508,
+      "loss/crossentropy": 2.516406524181366,
+      "loss/hidden": 3.35,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35599096268415453,
+      "loss/reg": 0.0,
+      "step": 5900
+    },
+    {
+      "epoch": 0.03888157894736842,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.19685872395833334,
+      "learning_rate": 0.0001,
+      "loss": 3.3701,
+      "loss/crossentropy": 2.118768775463104,
+      "loss/hidden": 3.234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3433256149291992,
+      "loss/reg": 0.0,
+      "step": 5910
+    },
+    {
+      "epoch": 0.03894736842105263,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.10144856770833334,
+      "learning_rate": 0.0001,
+      "loss": 3.4698,
+      "loss/crossentropy": 2.4102617263793946,
+      "loss/hidden": 3.059375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2903832048177719,
+      "loss/reg": 0.0,
+      "step": 5920
+    },
+    {
+      "epoch": 0.03901315789473684,
+      "grad_norm": 3.25,
+      "grad_norm_var": 0.05016276041666667,
+      "learning_rate": 0.0001,
+      "loss": 3.3435,
+      "loss/crossentropy": 2.4927979469299317,
+      "loss/hidden": 3.046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2977334216237068,
+      "loss/reg": 0.0,
+      "step": 5930
+    },
+    {
+      "epoch": 0.03907894736842105,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.2647043863932292,
+      "learning_rate": 0.0001,
+      "loss": 3.4123,
+      "loss/crossentropy": 2.163569325208664,
+      "loss/hidden": 3.1515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27049526423215864,
+      "loss/reg": 0.0,
+      "step": 5940
+    },
+    {
+      "epoch": 0.03914473684210526,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.1790679931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.4968,
+      "loss/crossentropy": 2.314089775085449,
+      "loss/hidden": 3.290625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2804916575551033,
+      "loss/reg": 0.0,
+      "step": 5950
+    },
+    {
+      "epoch": 0.03921052631578947,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.10930989583333334,
+      "learning_rate": 0.0001,
+      "loss": 3.4204,
+      "loss/crossentropy": 2.5095210552215574,
+      "loss/hidden": 3.1890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34905528128147123,
+      "loss/reg": 0.0,
+      "step": 5960
+    },
+    {
+      "epoch": 0.03927631578947369,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.024348958333333334,
+      "learning_rate": 0.0001,
+      "loss": 3.3368,
+      "loss/crossentropy": 2.3903687596321106,
+      "loss/hidden": 3.228125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36534676551818845,
+      "loss/reg": 0.0,
+      "step": 5970
+    },
+    {
+      "epoch": 0.039342105263157894,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.04106343587239583,
+      "learning_rate": 0.0001,
+      "loss": 3.3975,
+      "loss/crossentropy": 2.485898661613464,
+      "loss/hidden": 2.946875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28061621338129045,
+      "loss/reg": 0.0,
+      "step": 5980
+    },
+    {
+      "epoch": 0.03940789473684211,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.0488677978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.3803,
+      "loss/crossentropy": 2.4124781847000123,
+      "loss/hidden": 3.009375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30112328827381135,
+      "loss/reg": 0.0,
+      "step": 5990
+    },
+    {
+      "epoch": 0.039473684210526314,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.44810791015625,
+      "learning_rate": 0.0001,
+      "loss": 3.472,
+      "loss/crossentropy": 2.4459670901298525,
+      "loss/hidden": 2.9671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2799991726875305,
+      "loss/reg": 0.0,
+      "step": 6000
+    },
+    {
+      "epoch": 0.03953947368421053,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.1685546875,
+      "learning_rate": 0.0001,
+      "loss": 3.4322,
+      "loss/crossentropy": 2.641672468185425,
+      "loss/hidden": 3.0203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31368287801742556,
+      "loss/reg": 0.0,
+      "step": 6010
+    },
+    {
+      "epoch": 0.039605263157894734,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.03518473307291667,
+      "learning_rate": 0.0001,
+      "loss": 3.4325,
+      "loss/crossentropy": 2.2493654131889342,
+      "loss/hidden": 3.134375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29540172666311265,
+      "loss/reg": 0.0,
+      "step": 6020
+    },
+    {
+      "epoch": 0.03967105263157895,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.04810282389322917,
+      "learning_rate": 0.0001,
+      "loss": 3.408,
+      "loss/crossentropy": 2.5626556158065794,
+      "loss/hidden": 3.178125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3445401757955551,
+      "loss/reg": 0.0,
+      "step": 6030
+    },
+    {
+      "epoch": 0.03973684210526316,
+      "grad_norm": 3.375,
+      "grad_norm_var": 0.11850484212239583,
+      "learning_rate": 0.0001,
+      "loss": 3.3697,
+      "loss/crossentropy": 2.2249147415161135,
+      "loss/hidden": 3.165625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3320572040975094,
+      "loss/reg": 0.0,
+      "step": 6040
+    },
+    {
+      "epoch": 0.03980263157894737,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.1265045166015625,
+      "learning_rate": 0.0001,
+      "loss": 3.3726,
+      "loss/crossentropy": 2.479216980934143,
+      "loss/hidden": 3.021875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3085744693875313,
+      "loss/reg": 0.0,
+      "step": 6050
+    },
+    {
+      "epoch": 0.03986842105263158,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.2034088134765625,
+      "learning_rate": 0.0001,
+      "loss": 3.4346,
+      "loss/crossentropy": 2.3974440932273864,
+      "loss/hidden": 3.259375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4024490460753441,
+      "loss/reg": 0.0,
+      "step": 6060
+    },
+    {
+      "epoch": 0.03993421052631579,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.19241434733072918,
+      "learning_rate": 0.0001,
+      "loss": 3.3911,
+      "loss/crossentropy": 2.3311298370361326,
+      "loss/hidden": 3.00625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2968046382069588,
+      "loss/reg": 0.0,
+      "step": 6070
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.060445149739583336,
+      "learning_rate": 0.0001,
+      "loss": 3.5175,
+      "loss/crossentropy": 2.6493954181671144,
+      "loss/hidden": 3.2828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3133848324418068,
+      "loss/reg": 0.0,
+      "step": 6080
+    },
+    {
+      "epoch": 0.04006578947368421,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.2754140218098958,
+      "learning_rate": 0.0001,
+      "loss": 3.3932,
+      "loss/crossentropy": 2.4315222024917604,
+      "loss/hidden": 3.159375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3378627926111221,
+      "loss/reg": 0.0,
+      "step": 6090
+    },
+    {
+      "epoch": 0.04013157894736842,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.04810282389322917,
+      "learning_rate": 0.0001,
+      "loss": 3.3698,
+      "loss/crossentropy": 2.421731984615326,
+      "loss/hidden": 3.315625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3809585988521576,
+      "loss/reg": 0.0,
+      "step": 6100
+    },
+    {
+      "epoch": 0.040197368421052634,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.0509429931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.3839,
+      "loss/crossentropy": 2.2816696763038635,
+      "loss/hidden": 3.0390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3188880756497383,
+      "loss/reg": 0.0,
+      "step": 6110
+    },
+    {
+      "epoch": 0.04026315789473684,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.26266276041666664,
+      "learning_rate": 0.0001,
+      "loss": 3.4852,
+      "loss/crossentropy": 2.4251498103141786,
+      "loss/hidden": 2.9875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3131637305021286,
+      "loss/reg": 0.0,
+      "step": 6120
+    },
+    {
+      "epoch": 0.040328947368421054,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.3377919514973958,
+      "learning_rate": 0.0001,
+      "loss": 3.4191,
+      "loss/crossentropy": 2.321718716621399,
+      "loss/hidden": 3.028125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28195892125368116,
+      "loss/reg": 0.0,
+      "step": 6130
+    },
+    {
+      "epoch": 0.04039473684210526,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.0685943603515625,
+      "learning_rate": 0.0001,
+      "loss": 3.4412,
+      "loss/crossentropy": 2.658420753479004,
+      "loss/hidden": 3.125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3355125278234482,
+      "loss/reg": 0.0,
+      "step": 6140
+    },
+    {
+      "epoch": 0.040460526315789475,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.059305826822916664,
+      "learning_rate": 0.0001,
+      "loss": 3.3292,
+      "loss/crossentropy": 2.5203867316246034,
+      "loss/hidden": 2.9859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3060264021158218,
+      "loss/reg": 0.0,
+      "step": 6150
+    },
+    {
+      "epoch": 0.04052631578947368,
+      "grad_norm": 3.078125,
+      "grad_norm_var": 0.052302042643229164,
+      "learning_rate": 0.0001,
+      "loss": 3.3844,
+      "loss/crossentropy": 2.53420695066452,
+      "loss/hidden": 3.178125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3124456375837326,
+      "loss/reg": 0.0,
+      "step": 6160
+    },
+    {
+      "epoch": 0.040592105263157895,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.049779256184895836,
+      "learning_rate": 0.0001,
+      "loss": 3.3634,
+      "loss/crossentropy": 2.5132151365280153,
+      "loss/hidden": 3.128125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35312571823596955,
+      "loss/reg": 0.0,
+      "step": 6170
+    },
+    {
+      "epoch": 0.04065789473684211,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.16035054524739584,
+      "learning_rate": 0.0001,
+      "loss": 3.4426,
+      "loss/crossentropy": 2.3004459500312806,
+      "loss/hidden": 3.1703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28007449954748154,
+      "loss/reg": 0.0,
+      "step": 6180
+    },
+    {
+      "epoch": 0.040723684210526315,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.1358306884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.3068,
+      "loss/crossentropy": 2.373980039358139,
+      "loss/hidden": 3.015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28091391175985336,
+      "loss/reg": 0.0,
+      "step": 6190
+    },
+    {
+      "epoch": 0.04078947368421053,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.08801981608072916,
+      "learning_rate": 0.0001,
+      "loss": 3.4006,
+      "loss/crossentropy": 2.2023098945617674,
+      "loss/hidden": 3.00625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2726595625281334,
+      "loss/reg": 0.0,
+      "step": 6200
+    },
+    {
+      "epoch": 0.040855263157894735,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 3.6811187744140623,
+      "learning_rate": 0.0001,
+      "loss": 3.6073,
+      "loss/crossentropy": 2.382032370567322,
+      "loss/hidden": 3.6921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4323269993066788,
+      "loss/reg": 0.0,
+      "step": 6210
+    },
+    {
+      "epoch": 0.04092105263157895,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 3.804295857747396,
+      "learning_rate": 0.0001,
+      "loss": 3.3627,
+      "loss/crossentropy": 2.569228994846344,
+      "loss/hidden": 3.0453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31191317439079286,
+      "loss/reg": 0.0,
+      "step": 6220
+    },
+    {
+      "epoch": 0.040986842105263155,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.08497721354166667,
+      "learning_rate": 0.0001,
+      "loss": 3.3414,
+      "loss/crossentropy": 2.5153043985366823,
+      "loss/hidden": 2.9609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30689269602298735,
+      "loss/reg": 0.0,
+      "step": 6230
+    },
+    {
+      "epoch": 0.04105263157894737,
+      "grad_norm": 2.375,
+      "grad_norm_var": 1.2814849853515624,
+      "learning_rate": 0.0001,
+      "loss": 3.4386,
+      "loss/crossentropy": 2.283157765865326,
+      "loss/hidden": 3.1203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3153227433562279,
+      "loss/reg": 0.0,
+      "step": 6240
+    },
+    {
+      "epoch": 0.04111842105263158,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 5.054263305664063,
+      "learning_rate": 0.0001,
+      "loss": 3.4379,
+      "loss/crossentropy": 2.585819673538208,
+      "loss/hidden": 2.9640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2913993000984192,
+      "loss/reg": 0.0,
+      "step": 6250
+    },
+    {
+      "epoch": 0.04118421052631579,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.0501373291015625,
+      "learning_rate": 0.0001,
+      "loss": 3.3357,
+      "loss/crossentropy": 2.453801620006561,
+      "loss/hidden": 3.0203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3249350532889366,
+      "loss/reg": 0.0,
+      "step": 6260
+    },
+    {
+      "epoch": 0.04125,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.0405181884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.3525,
+      "loss/crossentropy": 2.4949014663696287,
+      "loss/hidden": 3.00625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2902356445789337,
+      "loss/reg": 0.0,
+      "step": 6270
+    },
+    {
+      "epoch": 0.04131578947368421,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 1.6241770426432292,
+      "learning_rate": 0.0001,
+      "loss": 3.4444,
+      "loss/crossentropy": 2.029393529891968,
+      "loss/hidden": 3.1125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30831936225295065,
+      "loss/reg": 0.0,
+      "step": 6280
+    },
+    {
+      "epoch": 0.04138157894736842,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.0864654541015625,
+      "learning_rate": 0.0001,
+      "loss": 3.3944,
+      "loss/crossentropy": 3.0075352430343627,
+      "loss/hidden": 3.2390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4476942718029022,
+      "loss/reg": 0.0,
+      "step": 6290
+    },
+    {
+      "epoch": 0.04144736842105263,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.15415751139322917,
+      "learning_rate": 0.0001,
+      "loss": 3.4091,
+      "loss/crossentropy": 2.36070739030838,
+      "loss/hidden": 2.9109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26602610796689985,
+      "loss/reg": 0.0,
+      "step": 6300
+    },
+    {
+      "epoch": 0.04151315789473684,
+      "grad_norm": 3.0,
+      "grad_norm_var": 0.18596903483072916,
+      "learning_rate": 0.0001,
+      "loss": 3.4135,
+      "loss/crossentropy": 2.186020624637604,
+      "loss/hidden": 3.0203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30242343842983244,
+      "loss/reg": 0.0,
+      "step": 6310
+    },
+    {
+      "epoch": 0.041578947368421056,
+      "grad_norm": 3.34375,
+      "grad_norm_var": 0.06327718098958333,
+      "learning_rate": 0.0001,
+      "loss": 3.44,
+      "loss/crossentropy": 2.3751362919807435,
+      "loss/hidden": 2.99375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29995152205228803,
+      "loss/reg": 0.0,
+      "step": 6320
+    },
+    {
+      "epoch": 0.04164473684210526,
+      "grad_norm": 7.9375,
+      "grad_norm_var": 3.442041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.4985,
+      "loss/crossentropy": 2.2889585196971893,
+      "loss/hidden": 3.14375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5501813948154449,
+      "loss/reg": 0.0,
+      "step": 6330
+    },
+    {
+      "epoch": 0.041710526315789476,
+      "grad_norm": 2.875,
+      "grad_norm_var": 3.8211822509765625,
+      "learning_rate": 0.0001,
+      "loss": 3.4688,
+      "loss/crossentropy": 2.378446078300476,
+      "loss/hidden": 3.0609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30127355754375457,
+      "loss/reg": 0.0,
+      "step": 6340
+    },
+    {
+      "epoch": 0.04177631578947368,
+      "grad_norm": 3.109375,
+      "grad_norm_var": 1.1990793863932292,
+      "learning_rate": 0.0001,
+      "loss": 3.3813,
+      "loss/crossentropy": 2.444005084037781,
+      "loss/hidden": 3.209375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3470224469900131,
+      "loss/reg": 0.0,
+      "step": 6350
+    },
+    {
+      "epoch": 0.041842105263157896,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.7403065999348958,
+      "learning_rate": 0.0001,
+      "loss": 3.4,
+      "loss/crossentropy": 2.026668357849121,
+      "loss/hidden": 2.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2548553004860878,
+      "loss/reg": 0.0,
+      "step": 6360
+    },
+    {
+      "epoch": 0.0419078947368421,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.3556925455729167,
+      "learning_rate": 0.0001,
+      "loss": 3.4591,
+      "loss/crossentropy": 2.331345629692078,
+      "loss/hidden": 2.9796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26911164075136185,
+      "loss/reg": 0.0,
+      "step": 6370
+    },
+    {
+      "epoch": 0.041973684210526316,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.27215169270833334,
+      "learning_rate": 0.0001,
+      "loss": 3.4011,
+      "loss/crossentropy": 2.2243176221847536,
+      "loss/hidden": 3.0875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30012439042329786,
+      "loss/reg": 0.0,
+      "step": 6380
+    },
+    {
+      "epoch": 0.04203947368421053,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.0267974853515625,
+      "learning_rate": 0.0001,
+      "loss": 3.3476,
+      "loss/crossentropy": 2.284479832649231,
+      "loss/hidden": 3.034375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3452886208891869,
+      "loss/reg": 0.0,
+      "step": 6390
+    },
+    {
+      "epoch": 0.042105263157894736,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.18044331868489583,
+      "learning_rate": 0.0001,
+      "loss": 3.4206,
+      "loss/crossentropy": 2.2036523103713987,
+      "loss/hidden": 3.0515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2783824667334557,
+      "loss/reg": 0.0,
+      "step": 6400
+    },
+    {
+      "epoch": 0.04217105263157895,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 4.91226806640625,
+      "learning_rate": 0.0001,
+      "loss": 3.4892,
+      "loss/crossentropy": 2.2305456399917603,
+      "loss/hidden": 3.125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29731594026088715,
+      "loss/reg": 0.0,
+      "step": 6410
+    },
+    {
+      "epoch": 0.042236842105263156,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.8588216145833333,
+      "learning_rate": 0.0001,
+      "loss": 3.3759,
+      "loss/crossentropy": 2.186525213718414,
+      "loss/hidden": 3.090625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31351439356803895,
+      "loss/reg": 0.0,
+      "step": 6420
+    },
+    {
+      "epoch": 0.04230263157894737,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 1.168097941080729,
+      "learning_rate": 0.0001,
+      "loss": 3.4452,
+      "loss/crossentropy": 2.3028628826141357,
+      "loss/hidden": 3.29375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40216329991817473,
+      "loss/reg": 0.0,
+      "step": 6430
+    },
+    {
+      "epoch": 0.042368421052631576,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.0708160400390625,
+      "learning_rate": 0.0001,
+      "loss": 3.362,
+      "loss/crossentropy": 2.5469772100448607,
+      "loss/hidden": 3.10625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3354289785027504,
+      "loss/reg": 0.0,
+      "step": 6440
+    },
+    {
+      "epoch": 0.04243421052631579,
+      "grad_norm": 3.4375,
+      "grad_norm_var": 0.09075419108072917,
+      "learning_rate": 0.0001,
+      "loss": 3.366,
+      "loss/crossentropy": 2.3079045534133913,
+      "loss/hidden": 2.9953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2784201934933662,
+      "loss/reg": 0.0,
+      "step": 6450
+    },
+    {
+      "epoch": 0.0425,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.10932515462239584,
+      "learning_rate": 0.0001,
+      "loss": 3.4156,
+      "loss/crossentropy": 2.325330352783203,
+      "loss/hidden": 3.1828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3098024681210518,
+      "loss/reg": 0.0,
+      "step": 6460
+    },
+    {
+      "epoch": 0.04256578947368421,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.07423502604166667,
+      "learning_rate": 0.0001,
+      "loss": 3.3605,
+      "loss/crossentropy": 2.4809486865997314,
+      "loss/hidden": 3.146875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.336503566801548,
+      "loss/reg": 0.0,
+      "step": 6470
+    },
+    {
+      "epoch": 0.04263157894736842,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.06294657389322916,
+      "learning_rate": 0.0001,
+      "loss": 3.4219,
+      "loss/crossentropy": 2.357036566734314,
+      "loss/hidden": 3.16875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31517077386379244,
+      "loss/reg": 0.0,
+      "step": 6480
+    },
+    {
+      "epoch": 0.04269736842105263,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.044188435872395834,
+      "learning_rate": 0.0001,
+      "loss": 3.2698,
+      "loss/crossentropy": 2.3011206150054933,
+      "loss/hidden": 3.0296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3023343622684479,
+      "loss/reg": 0.0,
+      "step": 6490
+    },
+    {
+      "epoch": 0.04276315789473684,
+      "grad_norm": 3.75,
+      "grad_norm_var": 0.191259765625,
+      "learning_rate": 0.0001,
+      "loss": 3.3991,
+      "loss/crossentropy": 2.135625755786896,
+      "loss/hidden": 3.1328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30503824055194856,
+      "loss/reg": 0.0,
+      "step": 6500
+    },
+    {
+      "epoch": 0.04282894736842105,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.16857096354166667,
+      "learning_rate": 0.0001,
+      "loss": 3.317,
+      "loss/crossentropy": 2.2963179469108583,
+      "loss/hidden": 3.090625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30365400537848475,
+      "loss/reg": 0.0,
+      "step": 6510
+    },
+    {
+      "epoch": 0.04289473684210526,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.0795562744140625,
+      "learning_rate": 0.0001,
+      "loss": 3.3769,
+      "loss/crossentropy": 2.463003098964691,
+      "loss/hidden": 2.8953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2864454731345177,
+      "loss/reg": 0.0,
+      "step": 6520
+    },
+    {
+      "epoch": 0.04296052631578948,
+      "grad_norm": 3.28125,
+      "grad_norm_var": 0.1319244384765625,
+      "learning_rate": 0.0001,
+      "loss": 3.4418,
+      "loss/crossentropy": 2.4336194515228273,
+      "loss/hidden": 3.1328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2907303601503372,
+      "loss/reg": 0.0,
+      "step": 6530
+    },
+    {
+      "epoch": 0.04302631578947368,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.18455403645833332,
+      "learning_rate": 0.0001,
+      "loss": 3.5029,
+      "loss/crossentropy": 2.4530020356178284,
+      "loss/hidden": 3.0484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3084172964096069,
+      "loss/reg": 0.0,
+      "step": 6540
+    },
+    {
+      "epoch": 0.0430921052631579,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.0909332275390625,
+      "learning_rate": 0.0001,
+      "loss": 3.4113,
+      "loss/crossentropy": 2.28972727060318,
+      "loss/hidden": 3.09375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29103828966617584,
+      "loss/reg": 0.0,
+      "step": 6550
+    },
+    {
+      "epoch": 0.0431578947368421,
+      "grad_norm": 3.15625,
+      "grad_norm_var": 0.09442952473958334,
+      "learning_rate": 0.0001,
+      "loss": 3.3842,
+      "loss/crossentropy": 2.5410515666007996,
+      "loss/hidden": 2.9390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.262499050796032,
+      "loss/reg": 0.0,
+      "step": 6560
+    },
+    {
+      "epoch": 0.04322368421052632,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.5502105712890625,
+      "learning_rate": 0.0001,
+      "loss": 3.3708,
+      "loss/crossentropy": 2.460482358932495,
+      "loss/hidden": 3.040625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2910577103495598,
+      "loss/reg": 0.0,
+      "step": 6570
+    },
+    {
+      "epoch": 0.043289473684210523,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.5198527018229167,
+      "learning_rate": 0.0001,
+      "loss": 3.2716,
+      "loss/crossentropy": 2.264538216590881,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26389954835176466,
+      "loss/reg": 0.0,
+      "step": 6580
+    },
+    {
+      "epoch": 0.04335526315789474,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.58785400390625,
+      "learning_rate": 0.0001,
+      "loss": 3.4431,
+      "loss/crossentropy": 2.705476760864258,
+      "loss/hidden": 3.0921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3710309460759163,
+      "loss/reg": 0.0,
+      "step": 6590
+    },
+    {
+      "epoch": 0.04342105263157895,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.7066721598307292,
+      "learning_rate": 0.0001,
+      "loss": 3.3844,
+      "loss/crossentropy": 2.3109049081802366,
+      "loss/hidden": 3.0921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3492768794298172,
+      "loss/reg": 0.0,
+      "step": 6600
+    },
+    {
+      "epoch": 0.04348684210526316,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 2.2473958333333335,
+      "learning_rate": 0.0001,
+      "loss": 3.5366,
+      "loss/crossentropy": 2.110491228103638,
+      "loss/hidden": 3.2296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3138969212770462,
+      "loss/reg": 0.0,
+      "step": 6610
+    },
+    {
+      "epoch": 0.04355263157894737,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 1.2563222249348958,
+      "learning_rate": 0.0001,
+      "loss": 3.3583,
+      "loss/crossentropy": 2.5283448338508605,
+      "loss/hidden": 2.9859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2928563803434372,
+      "loss/reg": 0.0,
+      "step": 6620
+    },
+    {
+      "epoch": 0.04361842105263158,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 1.2412261962890625,
+      "learning_rate": 0.0001,
+      "loss": 3.3885,
+      "loss/crossentropy": 2.579999303817749,
+      "loss/hidden": 3.4296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4414908319711685,
+      "loss/reg": 0.0,
+      "step": 6630
+    },
+    {
+      "epoch": 0.04368421052631579,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.11516825358072917,
+      "learning_rate": 0.0001,
+      "loss": 3.3022,
+      "loss/crossentropy": 2.413753032684326,
+      "loss/hidden": 2.9546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28862773478031156,
+      "loss/reg": 0.0,
+      "step": 6640
+    },
+    {
+      "epoch": 0.04375,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.28503316243489585,
+      "learning_rate": 0.0001,
+      "loss": 3.4009,
+      "loss/crossentropy": 2.4203495264053343,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2926980495452881,
+      "loss/reg": 0.0,
+      "step": 6650
+    },
+    {
+      "epoch": 0.04381578947368421,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.07642822265625,
+      "learning_rate": 0.0001,
+      "loss": 3.4092,
+      "loss/crossentropy": 2.3588085770606995,
+      "loss/hidden": 3.2,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32405087500810625,
+      "loss/reg": 0.0,
+      "step": 6660
+    },
+    {
+      "epoch": 0.043881578947368424,
+      "grad_norm": 3.53125,
+      "grad_norm_var": 0.45767822265625,
+      "learning_rate": 0.0001,
+      "loss": 3.4143,
+      "loss/crossentropy": 2.512442636489868,
+      "loss/hidden": 3.0671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2854305922985077,
+      "loss/reg": 0.0,
+      "step": 6670
+    },
+    {
+      "epoch": 0.04394736842105263,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.3211873372395833,
+      "learning_rate": 0.0001,
+      "loss": 3.3297,
+      "loss/crossentropy": 2.5028780698776245,
+      "loss/hidden": 3.2421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3471944749355316,
+      "loss/reg": 0.0,
+      "step": 6680
+    },
+    {
+      "epoch": 0.044013157894736844,
+      "grad_norm": 4.0625,
+      "grad_norm_var": 0.8257720947265625,
+      "learning_rate": 0.0001,
+      "loss": 3.4063,
+      "loss/crossentropy": 2.2031276702880858,
+      "loss/hidden": 3.0546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25473351776599884,
+      "loss/reg": 0.0,
+      "step": 6690
+    },
+    {
+      "epoch": 0.04407894736842105,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.22024637858072918,
+      "learning_rate": 0.0001,
+      "loss": 3.4035,
+      "loss/crossentropy": 2.2040748953819276,
+      "loss/hidden": 3.434375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3361481264233589,
+      "loss/reg": 0.0,
+      "step": 6700
+    },
+    {
+      "epoch": 0.044144736842105264,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.1261383056640625,
+      "learning_rate": 0.0001,
+      "loss": 3.4225,
+      "loss/crossentropy": 2.390383541584015,
+      "loss/hidden": 3.01875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2891513243317604,
+      "loss/reg": 0.0,
+      "step": 6710
+    },
+    {
+      "epoch": 0.04421052631578947,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.0734375,
+      "learning_rate": 0.0001,
+      "loss": 3.4363,
+      "loss/crossentropy": 2.392831575870514,
+      "loss/hidden": 3.125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3119692116975784,
+      "loss/reg": 0.0,
+      "step": 6720
+    },
+    {
+      "epoch": 0.044276315789473684,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.12464192708333334,
+      "learning_rate": 0.0001,
+      "loss": 3.3959,
+      "loss/crossentropy": 2.426222395896912,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2563284829258919,
+      "loss/reg": 0.0,
+      "step": 6730
+    },
+    {
+      "epoch": 0.0443421052631579,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 3.1175201416015623,
+      "learning_rate": 0.0001,
+      "loss": 3.4427,
+      "loss/crossentropy": 2.408197546005249,
+      "loss/hidden": 3.0109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3056713670492172,
+      "loss/reg": 0.0,
+      "step": 6740
+    },
+    {
+      "epoch": 0.044407894736842105,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 2.6096099853515624,
+      "learning_rate": 0.0001,
+      "loss": 3.3365,
+      "loss/crossentropy": 2.5353691220283507,
+      "loss/hidden": 3.0046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30413212031126025,
+      "loss/reg": 0.0,
+      "step": 6750
+    },
+    {
+      "epoch": 0.04447368421052632,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.233740234375,
+      "learning_rate": 0.0001,
+      "loss": 3.371,
+      "loss/crossentropy": 2.495334494113922,
+      "loss/hidden": 3.0640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3251196876168251,
+      "loss/reg": 0.0,
+      "step": 6760
+    },
+    {
+      "epoch": 0.044539473684210525,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.11097005208333334,
+      "learning_rate": 0.0001,
+      "loss": 3.3306,
+      "loss/crossentropy": 2.3767608165740968,
+      "loss/hidden": 3.0796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27886694818735125,
+      "loss/reg": 0.0,
+      "step": 6770
+    },
+    {
+      "epoch": 0.04460526315789474,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.17870992024739582,
+      "learning_rate": 0.0001,
+      "loss": 3.4315,
+      "loss/crossentropy": 2.5281002640724184,
+      "loss/hidden": 2.9984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31765572130680086,
+      "loss/reg": 0.0,
+      "step": 6780
+    },
+    {
+      "epoch": 0.044671052631578945,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 3.0139973958333335,
+      "learning_rate": 0.0001,
+      "loss": 3.4717,
+      "loss/crossentropy": 2.362587594985962,
+      "loss/hidden": 3.1125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3029619336128235,
+      "loss/reg": 0.0,
+      "step": 6790
+    },
+    {
+      "epoch": 0.04473684210526316,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 7.6724192301432295,
+      "learning_rate": 0.0001,
+      "loss": 3.4302,
+      "loss/crossentropy": 2.4665863275527955,
+      "loss/hidden": 2.984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2934702351689339,
+      "loss/reg": 0.0,
+      "step": 6800
+    },
+    {
+      "epoch": 0.04480263157894737,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 7.714127604166666,
+      "learning_rate": 0.0001,
+      "loss": 3.4034,
+      "loss/crossentropy": 2.1714313626289368,
+      "loss/hidden": 3.1328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3513699471950531,
+      "loss/reg": 0.0,
+      "step": 6810
+    },
+    {
+      "epoch": 0.04486842105263158,
+      "grad_norm": 3.953125,
+      "grad_norm_var": 0.17561848958333334,
+      "learning_rate": 0.0001,
+      "loss": 3.3193,
+      "loss/crossentropy": 2.145359480381012,
+      "loss/hidden": 3.1015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.299596332013607,
+      "loss/reg": 0.0,
+      "step": 6820
+    },
+    {
+      "epoch": 0.04493421052631579,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.20790913899739583,
+      "learning_rate": 0.0001,
+      "loss": 3.343,
+      "loss/crossentropy": 2.7453124046325685,
+      "loss/hidden": 3.0921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36230285465717316,
+      "loss/reg": 0.0,
+      "step": 6830
+    },
+    {
+      "epoch": 0.045,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 2.236214192708333,
+      "learning_rate": 0.0001,
+      "loss": 3.4372,
+      "loss/crossentropy": 2.256898009777069,
+      "loss/hidden": 3.0546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2946140691637993,
+      "loss/reg": 0.0,
+      "step": 6840
+    },
+    {
+      "epoch": 0.04506578947368421,
+      "grad_norm": 2.375,
+      "grad_norm_var": 2.391950480143229,
+      "learning_rate": 0.0001,
+      "loss": 3.3701,
+      "loss/crossentropy": 2.6210601568222045,
+      "loss/hidden": 3.1015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3517232984304428,
+      "loss/reg": 0.0,
+      "step": 6850
+    },
+    {
+      "epoch": 0.04513157894736842,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 1.015998331705729,
+      "learning_rate": 0.0001,
+      "loss": 3.3998,
+      "loss/crossentropy": 2.387173318862915,
+      "loss/hidden": 3.3296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3826398134231567,
+      "loss/reg": 0.0,
+      "step": 6860
+    },
+    {
+      "epoch": 0.04519736842105263,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.44384663899739585,
+      "learning_rate": 0.0001,
+      "loss": 3.3858,
+      "loss/crossentropy": 2.4001118540763855,
+      "loss/hidden": 2.96875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27424332648515704,
+      "loss/reg": 0.0,
+      "step": 6870
+    },
+    {
+      "epoch": 0.045263157894736845,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.06037495930989583,
+      "learning_rate": 0.0001,
+      "loss": 3.3752,
+      "loss/crossentropy": 2.1074828147888183,
+      "loss/hidden": 3.0234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2777694225311279,
+      "loss/reg": 0.0,
+      "step": 6880
+    },
+    {
+      "epoch": 0.04532894736842105,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.1735504150390625,
+      "learning_rate": 0.0001,
+      "loss": 3.3792,
+      "loss/crossentropy": 2.2711395502090452,
+      "loss/hidden": 3.1296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3001497104763985,
+      "loss/reg": 0.0,
+      "step": 6890
+    },
+    {
+      "epoch": 0.045394736842105265,
+      "grad_norm": 3.0,
+      "grad_norm_var": 1.8758951822916667,
+      "learning_rate": 0.0001,
+      "loss": 3.4311,
+      "loss/crossentropy": 2.2329100012779235,
+      "loss/hidden": 3.3421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3293987289071083,
+      "loss/reg": 0.0,
+      "step": 6900
+    },
+    {
+      "epoch": 0.04546052631578947,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 1.793781534830729,
+      "learning_rate": 0.0001,
+      "loss": 3.2852,
+      "loss/crossentropy": 2.3576371729373933,
+      "loss/hidden": 2.940625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27053930461406706,
+      "loss/reg": 0.0,
+      "step": 6910
+    },
+    {
+      "epoch": 0.045526315789473686,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 1.989207967122396,
+      "learning_rate": 0.0001,
+      "loss": 3.4411,
+      "loss/crossentropy": 2.532450318336487,
+      "loss/hidden": 2.9,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2638275146484375,
+      "loss/reg": 0.0,
+      "step": 6920
+    },
+    {
+      "epoch": 0.04559210526315789,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 2.184137980143229,
+      "learning_rate": 0.0001,
+      "loss": 3.4069,
+      "loss/crossentropy": 2.4933431267738344,
+      "loss/hidden": 3.2640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32281421422958373,
+      "loss/reg": 0.0,
+      "step": 6930
+    },
+    {
+      "epoch": 0.045657894736842106,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.05158589680989583,
+      "learning_rate": 0.0001,
+      "loss": 3.284,
+      "loss/crossentropy": 2.2752655148506165,
+      "loss/hidden": 2.9734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2645736649632454,
+      "loss/reg": 0.0,
+      "step": 6940
+    },
+    {
+      "epoch": 0.04572368421052632,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.0625152587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.3456,
+      "loss/crossentropy": 2.290147030353546,
+      "loss/hidden": 3.0609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27036611288785933,
+      "loss/reg": 0.0,
+      "step": 6950
+    },
+    {
+      "epoch": 0.045789473684210526,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 6.027144368489584,
+      "learning_rate": 0.0001,
+      "loss": 3.4116,
+      "loss/crossentropy": 2.359883761405945,
+      "loss/hidden": 3.0078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29680062681436536,
+      "loss/reg": 0.0,
+      "step": 6960
+    },
+    {
+      "epoch": 0.04585526315789474,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 0.5505208333333333,
+      "learning_rate": 0.0001,
+      "loss": 3.334,
+      "loss/crossentropy": 2.4557218074798586,
+      "loss/hidden": 3.0875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2946802690625191,
+      "loss/reg": 0.0,
+      "step": 6970
+    },
+    {
+      "epoch": 0.045921052631578946,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.50865478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.3166,
+      "loss/crossentropy": 2.2482771933078767,
+      "loss/hidden": 3.034375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25328404903411866,
+      "loss/reg": 0.0,
+      "step": 6980
+    },
+    {
+      "epoch": 0.04598684210526316,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.045685831705729166,
+      "learning_rate": 0.0001,
+      "loss": 3.4432,
+      "loss/crossentropy": 2.3823115646839144,
+      "loss/hidden": 2.9953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29391862004995345,
+      "loss/reg": 0.0,
+      "step": 6990
+    },
+    {
+      "epoch": 0.046052631578947366,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.04589436848958333,
+      "learning_rate": 0.0001,
+      "loss": 3.4052,
+      "loss/crossentropy": 2.67444326877594,
+      "loss/hidden": 3.3140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.336136220395565,
+      "loss/reg": 0.0,
+      "step": 7000
+    },
+    {
+      "epoch": 0.04611842105263158,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.03218994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.3293,
+      "loss/crossentropy": 2.4088356614112856,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26314016729593276,
+      "loss/reg": 0.0,
+      "step": 7010
+    },
+    {
+      "epoch": 0.04618421052631579,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.044266764322916666,
+      "learning_rate": 0.0001,
+      "loss": 3.3201,
+      "loss/crossentropy": 2.210337924957275,
+      "loss/hidden": 2.996875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28177270889282224,
+      "loss/reg": 0.0,
+      "step": 7020
+    },
+    {
+      "epoch": 0.04625,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 1.1310780843098958,
+      "learning_rate": 0.0001,
+      "loss": 3.4065,
+      "loss/crossentropy": 2.3901759028434753,
+      "loss/hidden": 3.1921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3315755516290665,
+      "loss/reg": 0.0,
+      "step": 7030
+    },
+    {
+      "epoch": 0.04631578947368421,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.27060139973958336,
+      "learning_rate": 0.0001,
+      "loss": 3.4377,
+      "loss/crossentropy": 2.3308543801307677,
+      "loss/hidden": 3.0828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27955446392297745,
+      "loss/reg": 0.0,
+      "step": 7040
+    },
+    {
+      "epoch": 0.04638157894736842,
+      "grad_norm": 3.953125,
+      "grad_norm_var": 0.25728759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.3076,
+      "loss/crossentropy": 2.3288169384002684,
+      "loss/hidden": 2.909375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25437864363193513,
+      "loss/reg": 0.0,
+      "step": 7050
+    },
+    {
+      "epoch": 0.04644736842105263,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.20709228515625,
+      "learning_rate": 0.0001,
+      "loss": 3.439,
+      "loss/crossentropy": 2.033195120096207,
+      "loss/hidden": 3.0140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2639622241258621,
+      "loss/reg": 0.0,
+      "step": 7060
+    },
+    {
+      "epoch": 0.04651315789473684,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.84127197265625,
+      "learning_rate": 0.0001,
+      "loss": 3.3119,
+      "loss/crossentropy": 2.366466200351715,
+      "loss/hidden": 2.946875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2680037707090378,
+      "loss/reg": 0.0,
+      "step": 7070
+    },
+    {
+      "epoch": 0.04657894736842105,
+      "grad_norm": 11.9375,
+      "grad_norm_var": 5.57564697265625,
+      "learning_rate": 0.0001,
+      "loss": 3.4068,
+      "loss/crossentropy": 2.3484351873397826,
+      "loss/hidden": 3.3828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3572035223245621,
+      "loss/reg": 0.0,
+      "step": 7080
+    },
+    {
+      "epoch": 0.04664473684210527,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 5.670992024739584,
+      "learning_rate": 0.0001,
+      "loss": 3.5356,
+      "loss/crossentropy": 2.491948664188385,
+      "loss/hidden": 3.1484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33623201847076417,
+      "loss/reg": 0.0,
+      "step": 7090
+    },
+    {
+      "epoch": 0.04671052631578947,
+      "grad_norm": 7.90625,
+      "grad_norm_var": 1.8282297770182292,
+      "learning_rate": 0.0001,
+      "loss": 3.4057,
+      "loss/crossentropy": 2.3702210783958435,
+      "loss/hidden": 3.203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3451476514339447,
+      "loss/reg": 0.0,
+      "step": 7100
+    },
+    {
+      "epoch": 0.04677631578947369,
+      "grad_norm": 2.5,
+      "grad_norm_var": 1.74947509765625,
+      "learning_rate": 0.0001,
+      "loss": 3.4173,
+      "loss/crossentropy": 2.372307813167572,
+      "loss/hidden": 3.103125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30449149161577227,
+      "loss/reg": 0.0,
+      "step": 7110
+    },
+    {
+      "epoch": 0.04684210526315789,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.1269683837890625,
+      "learning_rate": 0.0001,
+      "loss": 3.3266,
+      "loss/crossentropy": 2.4220902919769287,
+      "loss/hidden": 2.9890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29380183219909667,
+      "loss/reg": 0.0,
+      "step": 7120
+    },
+    {
+      "epoch": 0.04690789473684211,
+      "grad_norm": 2.375,
+      "grad_norm_var": 5.4926503499348955,
+      "learning_rate": 0.0001,
+      "loss": 3.4299,
+      "loss/crossentropy": 2.333963227272034,
+      "loss/hidden": 3.0890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32264130711555483,
+      "loss/reg": 0.0,
+      "step": 7130
+    },
+    {
+      "epoch": 0.04697368421052631,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 8.575126139322917,
+      "learning_rate": 0.0001,
+      "loss": 3.5129,
+      "loss/crossentropy": 2.447948896884918,
+      "loss/hidden": 3.0390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34722310602664946,
+      "loss/reg": 0.0,
+      "step": 7140
+    },
+    {
+      "epoch": 0.04703947368421053,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 4.295670572916666,
+      "learning_rate": 0.0001,
+      "loss": 3.4954,
+      "loss/crossentropy": 2.4875372767448427,
+      "loss/hidden": 3.1765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38640123009681704,
+      "loss/reg": 0.0,
+      "step": 7150
+    },
+    {
+      "epoch": 0.04710526315789473,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.09901936848958333,
+      "learning_rate": 0.0001,
+      "loss": 3.3592,
+      "loss/crossentropy": 2.4286764740943907,
+      "loss/hidden": 3.1515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3441846176981926,
+      "loss/reg": 0.0,
+      "step": 7160
+    },
+    {
+      "epoch": 0.04717105263157895,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.09706624348958333,
+      "learning_rate": 0.0001,
+      "loss": 3.4219,
+      "loss/crossentropy": 2.197064208984375,
+      "loss/hidden": 3.215625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30653059035539626,
+      "loss/reg": 0.0,
+      "step": 7170
+    },
+    {
+      "epoch": 0.04723684210526316,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.07274983723958334,
+      "learning_rate": 0.0001,
+      "loss": 3.4754,
+      "loss/crossentropy": 2.592367339134216,
+      "loss/hidden": 3.2078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30859925150871276,
+      "loss/reg": 0.0,
+      "step": 7180
+    },
+    {
+      "epoch": 0.04730263157894737,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.04854227701822917,
+      "learning_rate": 0.0001,
+      "loss": 3.335,
+      "loss/crossentropy": 2.4139750719070436,
+      "loss/hidden": 2.95,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26624749004840853,
+      "loss/reg": 0.0,
+      "step": 7190
+    },
+    {
+      "epoch": 0.04736842105263158,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.061579386393229164,
+      "learning_rate": 0.0001,
+      "loss": 3.4049,
+      "loss/crossentropy": 1.9312179803848266,
+      "loss/hidden": 2.9046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23240152448415757,
+      "loss/reg": 0.0,
+      "step": 7200
+    },
+    {
+      "epoch": 0.04743421052631579,
+      "grad_norm": 3.140625,
+      "grad_norm_var": 2.7487790626051414e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.5157,
+      "loss/crossentropy": 2.192197346687317,
+      "loss/hidden": 3.1421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28352061808109286,
+      "loss/reg": 0.0,
+      "step": 7210
+    },
+    {
+      "epoch": 0.0475,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 2.7487790627662506e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.3283,
+      "loss/crossentropy": 2.2706116318702696,
+      "loss/hidden": 3.0734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29702268838882445,
+      "loss/reg": 0.0,
+      "step": 7220
+    },
+    {
+      "epoch": 0.04756578947368421,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.05640869140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2627,
+      "loss/crossentropy": 2.354196774959564,
+      "loss/hidden": 3.0625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2917496845126152,
+      "loss/reg": 0.0,
+      "step": 7230
+    },
+    {
+      "epoch": 0.04763157894736842,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.048779296875,
+      "learning_rate": 0.0001,
+      "loss": 3.3302,
+      "loss/crossentropy": 2.4912230253219603,
+      "loss/hidden": 3.0265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3378173440694809,
+      "loss/reg": 0.0,
+      "step": 7240
+    },
+    {
+      "epoch": 0.047697368421052634,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.0933990478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.348,
+      "loss/crossentropy": 2.541353499889374,
+      "loss/hidden": 3.0359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3061401903629303,
+      "loss/reg": 0.0,
+      "step": 7250
+    },
+    {
+      "epoch": 0.04776315789473684,
+      "grad_norm": 4.15625,
+      "grad_norm_var": 0.45806884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.5007,
+      "loss/crossentropy": 2.4307178616523744,
+      "loss/hidden": 3.1921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3470552325248718,
+      "loss/reg": 0.0,
+      "step": 7260
+    },
+    {
+      "epoch": 0.047828947368421054,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.4556955973307292,
+      "learning_rate": 0.0001,
+      "loss": 3.3409,
+      "loss/crossentropy": 2.257239353656769,
+      "loss/hidden": 3.0546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29585833847522736,
+      "loss/reg": 0.0,
+      "step": 7270
+    },
+    {
+      "epoch": 0.04789473684210526,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.3855445861816406,
+      "learning_rate": 0.0001,
+      "loss": 3.2964,
+      "loss/crossentropy": 2.1451990723609926,
+      "loss/hidden": 2.9828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25767376720905305,
+      "loss/reg": 0.0,
+      "step": 7280
+    },
+    {
+      "epoch": 0.047960526315789474,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.26484349568684895,
+      "learning_rate": 0.0001,
+      "loss": 3.3283,
+      "loss/crossentropy": 2.2842231035232543,
+      "loss/hidden": 3.1171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33771214783191683,
+      "loss/reg": 0.0,
+      "step": 7290
+    },
+    {
+      "epoch": 0.04802631578947368,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.13931884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.2818,
+      "loss/crossentropy": 2.1290991365909577,
+      "loss/hidden": 2.8890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2695446103811264,
+      "loss/reg": 0.0,
+      "step": 7300
+    },
+    {
+      "epoch": 0.048092105263157894,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.0911773681640625,
+      "learning_rate": 0.0001,
+      "loss": 3.3833,
+      "loss/crossentropy": 2.149968445301056,
+      "loss/hidden": 2.9625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24086009413003923,
+      "loss/reg": 0.0,
+      "step": 7310
+    },
+    {
+      "epoch": 0.04815789473684211,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.07766011555989584,
+      "learning_rate": 0.0001,
+      "loss": 3.3503,
+      "loss/crossentropy": 2.519470489025116,
+      "loss/hidden": 2.890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25514234602451324,
+      "loss/reg": 0.0,
+      "step": 7320
+    },
+    {
+      "epoch": 0.048223684210526314,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.025614420572916668,
+      "learning_rate": 0.0001,
+      "loss": 3.3595,
+      "loss/crossentropy": 2.5252918124198915,
+      "loss/hidden": 3.0109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.281466007232666,
+      "loss/reg": 0.0,
+      "step": 7330
+    },
+    {
+      "epoch": 0.04828947368421053,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.37810872395833334,
+      "learning_rate": 0.0001,
+      "loss": 3.3791,
+      "loss/crossentropy": 2.0419042229652407,
+      "loss/hidden": 3.0515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2731199100613594,
+      "loss/reg": 0.0,
+      "step": 7340
+    },
+    {
+      "epoch": 0.048355263157894735,
+      "grad_norm": 3.421875,
+      "grad_norm_var": 0.19280192057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.3562,
+      "loss/crossentropy": 2.1462841510772703,
+      "loss/hidden": 3.04375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23967409282922744,
+      "loss/reg": 0.0,
+      "step": 7350
+    },
+    {
+      "epoch": 0.04842105263157895,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.13990478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.2953,
+      "loss/crossentropy": 2.4041597843170166,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26162110567092894,
+      "loss/reg": 0.0,
+      "step": 7360
+    },
+    {
+      "epoch": 0.048486842105263155,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.0317047119140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2731,
+      "loss/crossentropy": 2.61968252658844,
+      "loss/hidden": 2.9453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28658540844917296,
+      "loss/reg": 0.0,
+      "step": 7370
+    },
+    {
+      "epoch": 0.04855263157894737,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.0795562744140625,
+      "learning_rate": 0.0001,
+      "loss": 3.3537,
+      "loss/crossentropy": 2.330329430103302,
+      "loss/hidden": 3.0375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.311858968436718,
+      "loss/reg": 0.0,
+      "step": 7380
+    },
+    {
+      "epoch": 0.04861842105263158,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.06910400390625,
+      "learning_rate": 0.0001,
+      "loss": 3.3762,
+      "loss/crossentropy": 2.331431567668915,
+      "loss/hidden": 3.14375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3637159377336502,
+      "loss/reg": 0.0,
+      "step": 7390
+    },
+    {
+      "epoch": 0.04868421052631579,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.07679036458333334,
+      "learning_rate": 0.0001,
+      "loss": 3.3474,
+      "loss/crossentropy": 2.334563136100769,
+      "loss/hidden": 3.1203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3102908283472061,
+      "loss/reg": 0.0,
+      "step": 7400
+    },
+    {
+      "epoch": 0.04875,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.041727701822916664,
+      "learning_rate": 0.0001,
+      "loss": 3.2994,
+      "loss/crossentropy": 2.4575919032096865,
+      "loss/hidden": 3.1453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30418373495340345,
+      "loss/reg": 0.0,
+      "step": 7410
+    },
+    {
+      "epoch": 0.04881578947368421,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.95484619140625,
+      "learning_rate": 0.0001,
+      "loss": 3.3136,
+      "loss/crossentropy": 2.2615838646888733,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27002790868282317,
+      "loss/reg": 0.0,
+      "step": 7420
+    },
+    {
+      "epoch": 0.04888157894736842,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.05439046223958333,
+      "learning_rate": 0.0001,
+      "loss": 3.3084,
+      "loss/crossentropy": 2.2472564220428466,
+      "loss/hidden": 2.9703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28446437120437623,
+      "loss/reg": 0.0,
+      "step": 7430
+    },
+    {
+      "epoch": 0.04894736842105263,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.08689676920572917,
+      "learning_rate": 0.0001,
+      "loss": 3.3891,
+      "loss/crossentropy": 2.5821902751922607,
+      "loss/hidden": 3.21875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29424687922000886,
+      "loss/reg": 0.0,
+      "step": 7440
+    },
+    {
+      "epoch": 0.04901315789473684,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.09962565104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.4691,
+      "loss/crossentropy": 2.547790551185608,
+      "loss/hidden": 3.1125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3217499524354935,
+      "loss/reg": 0.0,
+      "step": 7450
+    },
+    {
+      "epoch": 0.049078947368421055,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.18788655598958334,
+      "learning_rate": 0.0001,
+      "loss": 3.3014,
+      "loss/crossentropy": 2.025312936306,
+      "loss/hidden": 3.0359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29601537734270095,
+      "loss/reg": 0.0,
+      "step": 7460
+    },
+    {
+      "epoch": 0.04914473684210526,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.15539957682291666,
+      "learning_rate": 0.0001,
+      "loss": 3.3516,
+      "loss/crossentropy": 2.1228564500808718,
+      "loss/hidden": 3.01875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2723393976688385,
+      "loss/reg": 0.0,
+      "step": 7470
+    },
+    {
+      "epoch": 0.049210526315789475,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.04488016764322917,
+      "learning_rate": 0.0001,
+      "loss": 3.4105,
+      "loss/crossentropy": 2.221798670291901,
+      "loss/hidden": 3.303125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3143211781978607,
+      "loss/reg": 0.0,
+      "step": 7480
+    },
+    {
+      "epoch": 0.04927631578947368,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.028938802083333333,
+      "learning_rate": 0.0001,
+      "loss": 3.4092,
+      "loss/crossentropy": 2.4991084337234497,
+      "loss/hidden": 3.24375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3354015931487083,
+      "loss/reg": 0.0,
+      "step": 7490
+    },
+    {
+      "epoch": 0.049342105263157895,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.04719136555989583,
+      "learning_rate": 0.0001,
+      "loss": 3.2643,
+      "loss/crossentropy": 2.3538936495780947,
+      "loss/hidden": 2.9890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25752398669719695,
+      "loss/reg": 0.0,
+      "step": 7500
+    },
+    {
+      "epoch": 0.0494078947368421,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.5662272135416667,
+      "learning_rate": 0.0001,
+      "loss": 3.3662,
+      "loss/crossentropy": 2.2441036820411684,
+      "loss/hidden": 2.9078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26160079389810564,
+      "loss/reg": 0.0,
+      "step": 7510
+    },
+    {
+      "epoch": 0.049473684210526316,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.5289703369140625,
+      "learning_rate": 0.0001,
+      "loss": 3.3149,
+      "loss/crossentropy": 2.544923734664917,
+      "loss/hidden": 3.015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2898894131183624,
+      "loss/reg": 0.0,
+      "step": 7520
+    },
+    {
+      "epoch": 0.04953947368421053,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.0388092041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.3611,
+      "loss/crossentropy": 2.5786613702774046,
+      "loss/hidden": 2.978125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2799048855900764,
+      "loss/reg": 0.0,
+      "step": 7530
+    },
+    {
+      "epoch": 0.049605263157894736,
+      "grad_norm": 4.625,
+      "grad_norm_var": 0.3494293212890625,
+      "learning_rate": 0.0001,
+      "loss": 3.3299,
+      "loss/crossentropy": 2.6309832334518433,
+      "loss/hidden": 2.9171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3181111514568329,
+      "loss/reg": 0.0,
+      "step": 7540
+    },
+    {
+      "epoch": 0.04967105263157895,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 0.32229715983072915,
+      "learning_rate": 0.0001,
+      "loss": 3.381,
+      "loss/crossentropy": 2.395383334159851,
+      "loss/hidden": 3.0484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29645184725522994,
+      "loss/reg": 0.0,
+      "step": 7550
+    },
+    {
+      "epoch": 0.049736842105263156,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.06101786295572917,
+      "learning_rate": 0.0001,
+      "loss": 3.3076,
+      "loss/crossentropy": 2.2212601780891417,
+      "loss/hidden": 3.0890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2814889296889305,
+      "loss/reg": 0.0,
+      "step": 7560
+    },
+    {
+      "epoch": 0.04980263157894737,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.5580800374348959,
+      "learning_rate": 0.0001,
+      "loss": 3.3547,
+      "loss/crossentropy": 2.340949076414108,
+      "loss/hidden": 3.2046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3048363208770752,
+      "loss/reg": 0.0,
+      "step": 7570
+    },
+    {
+      "epoch": 0.049868421052631576,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.5829498291015625,
+      "learning_rate": 0.0001,
+      "loss": 3.2873,
+      "loss/crossentropy": 2.434855592250824,
+      "loss/hidden": 3.1296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33333509862422944,
+      "loss/reg": 0.0,
+      "step": 7580
+    },
+    {
+      "epoch": 0.04993421052631579,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.07898661295572916,
+      "learning_rate": 0.0001,
+      "loss": 3.3901,
+      "loss/crossentropy": 2.2928370952606203,
+      "loss/hidden": 3.0484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30124022662639616,
+      "loss/reg": 0.0,
+      "step": 7590
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.04219462076822917,
+      "learning_rate": 0.0001,
+      "loss": 3.3048,
+      "loss/crossentropy": 2.0018354773521425,
+      "loss/hidden": 3.0796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25127379447221754,
+      "loss/reg": 0.0,
+      "step": 7600
+    },
+    {
+      "epoch": 0.05006578947368421,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.02652587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.3285,
+      "loss/crossentropy": 2.2610169410705567,
+      "loss/hidden": 3.0640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27540155351161955,
+      "loss/reg": 0.0,
+      "step": 7610
+    },
+    {
+      "epoch": 0.05013157894736842,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.0210357666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.2714,
+      "loss/crossentropy": 2.2387811303138734,
+      "loss/hidden": 2.9671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27260326892137526,
+      "loss/reg": 0.0,
+      "step": 7620
+    },
+    {
+      "epoch": 0.05019736842105263,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.25825093587239584,
+      "learning_rate": 0.0001,
+      "loss": 3.4396,
+      "loss/crossentropy": 2.5094308257102966,
+      "loss/hidden": 3.146875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3727137431502342,
+      "loss/reg": 0.0,
+      "step": 7630
+    },
+    {
+      "epoch": 0.05026315789473684,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.42695210774739584,
+      "learning_rate": 0.0001,
+      "loss": 3.4951,
+      "loss/crossentropy": 2.181921923160553,
+      "loss/hidden": 3.10625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29040979146957396,
+      "loss/reg": 0.0,
+      "step": 7640
+    },
+    {
+      "epoch": 0.05032894736842105,
+      "grad_norm": 3.5625,
+      "grad_norm_var": 0.47198893229166666,
+      "learning_rate": 0.0001,
+      "loss": 3.3254,
+      "loss/crossentropy": 2.4378631830215456,
+      "loss/hidden": 2.9984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32593746185302735,
+      "loss/reg": 0.0,
+      "step": 7650
+    },
+    {
+      "epoch": 0.05039473684210526,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.12525126139322917,
+      "learning_rate": 0.0001,
+      "loss": 3.2625,
+      "loss/crossentropy": 2.318000388145447,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.267861607670784,
+      "loss/reg": 0.0,
+      "step": 7660
+    },
+    {
+      "epoch": 0.050460526315789477,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.09370015462239584,
+      "learning_rate": 0.0001,
+      "loss": 3.4075,
+      "loss/crossentropy": 2.300287425518036,
+      "loss/hidden": 2.9609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.272597573697567,
+      "loss/reg": 0.0,
+      "step": 7670
+    },
+    {
+      "epoch": 0.05052631578947368,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.09342041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.2909,
+      "loss/crossentropy": 2.5380281090736387,
+      "loss/hidden": 2.9046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2639786213636398,
+      "loss/reg": 0.0,
+      "step": 7680
+    },
+    {
+      "epoch": 0.0505921052631579,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.165625,
+      "learning_rate": 0.0001,
+      "loss": 3.2847,
+      "loss/crossentropy": 2.0959218978881835,
+      "loss/hidden": 2.8453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2350485235452652,
+      "loss/reg": 0.0,
+      "step": 7690
+    },
+    {
+      "epoch": 0.0506578947368421,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.17685546875,
+      "learning_rate": 0.0001,
+      "loss": 3.3983,
+      "loss/crossentropy": 2.323216736316681,
+      "loss/hidden": 3.3,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3249870762228966,
+      "loss/reg": 0.0,
+      "step": 7700
+    },
+    {
+      "epoch": 0.05072368421052632,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.0990234375,
+      "learning_rate": 0.0001,
+      "loss": 3.3125,
+      "loss/crossentropy": 2.3656753659248353,
+      "loss/hidden": 2.99375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3182197526097298,
+      "loss/reg": 0.0,
+      "step": 7710
+    },
+    {
+      "epoch": 0.05078947368421052,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.19866129557291667,
+      "learning_rate": 0.0001,
+      "loss": 3.4009,
+      "loss/crossentropy": 2.5273406386375425,
+      "loss/hidden": 3.040625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3028098613023758,
+      "loss/reg": 0.0,
+      "step": 7720
+    },
+    {
+      "epoch": 0.05085526315789474,
+      "grad_norm": 3.234375,
+      "grad_norm_var": 0.20156148274739583,
+      "learning_rate": 0.0001,
+      "loss": 3.4345,
+      "loss/crossentropy": 2.391481709480286,
+      "loss/hidden": 3.175,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3816041976213455,
+      "loss/reg": 0.0,
+      "step": 7730
+    },
+    {
+      "epoch": 0.05092105263157895,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 0.08870035807291667,
+      "learning_rate": 0.0001,
+      "loss": 3.3338,
+      "loss/crossentropy": 2.4912365436553956,
+      "loss/hidden": 2.925,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2999194458127022,
+      "loss/reg": 0.0,
+      "step": 7740
+    },
+    {
+      "epoch": 0.05098684210526316,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.32692057291666665,
+      "learning_rate": 0.0001,
+      "loss": 3.3967,
+      "loss/crossentropy": 2.3413360595703123,
+      "loss/hidden": 3.225,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2852958709001541,
+      "loss/reg": 0.0,
+      "step": 7750
+    },
+    {
+      "epoch": 0.05105263157894737,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.32123921712239584,
+      "learning_rate": 0.0001,
+      "loss": 3.3178,
+      "loss/crossentropy": 2.6519938707351685,
+      "loss/hidden": 3.0609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30396163165569307,
+      "loss/reg": 0.0,
+      "step": 7760
+    },
+    {
+      "epoch": 0.05111842105263158,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.05095113118489583,
+      "learning_rate": 0.0001,
+      "loss": 3.3751,
+      "loss/crossentropy": 2.4069360971450804,
+      "loss/hidden": 2.884375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2661174476146698,
+      "loss/reg": 0.0,
+      "step": 7770
+    },
+    {
+      "epoch": 0.05118421052631579,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.12613016764322918,
+      "learning_rate": 0.0001,
+      "loss": 3.3737,
+      "loss/crossentropy": 2.3529985427856444,
+      "loss/hidden": 3.08125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2910769283771515,
+      "loss/reg": 0.0,
+      "step": 7780
+    },
+    {
+      "epoch": 0.05125,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.03486226399739583,
+      "learning_rate": 0.0001,
+      "loss": 3.3208,
+      "loss/crossentropy": 2.1253769397735596,
+      "loss/hidden": 2.9390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2470591977238655,
+      "loss/reg": 0.0,
+      "step": 7790
+    },
+    {
+      "epoch": 0.05131578947368421,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.5931925455729167,
+      "learning_rate": 0.0001,
+      "loss": 3.4704,
+      "loss/crossentropy": 2.2813072860240937,
+      "loss/hidden": 3.3578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3708792179822922,
+      "loss/reg": 0.0,
+      "step": 7800
+    },
+    {
+      "epoch": 0.051381578947368424,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.6090159098307292,
+      "learning_rate": 0.0001,
+      "loss": 3.3966,
+      "loss/crossentropy": 2.071128582954407,
+      "loss/hidden": 3.2015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.332095867395401,
+      "loss/reg": 0.0,
+      "step": 7810
+    },
+    {
+      "epoch": 0.05144736842105263,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.19628499348958334,
+      "learning_rate": 0.0001,
+      "loss": 3.3989,
+      "loss/crossentropy": 2.368660008907318,
+      "loss/hidden": 3.15625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28022406101226804,
+      "loss/reg": 0.0,
+      "step": 7820
+    },
+    {
+      "epoch": 0.051513157894736844,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 2.9096995035807294,
+      "learning_rate": 0.0001,
+      "loss": 3.3211,
+      "loss/crossentropy": 2.1510692477226256,
+      "loss/hidden": 2.890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2576268881559372,
+      "loss/reg": 0.0,
+      "step": 7830
+    },
+    {
+      "epoch": 0.05157894736842105,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 2.7149729410807293,
+      "learning_rate": 0.0001,
+      "loss": 3.4215,
+      "loss/crossentropy": 2.4896273493766783,
+      "loss/hidden": 3.4703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3669875577092171,
+      "loss/reg": 0.0,
+      "step": 7840
+    },
+    {
+      "epoch": 0.051644736842105264,
+      "grad_norm": 6.125,
+      "grad_norm_var": 0.9022450764973958,
+      "learning_rate": 0.0001,
+      "loss": 3.4161,
+      "loss/crossentropy": 2.0616636157035826,
+      "loss/hidden": 3.053125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2621200427412987,
+      "loss/reg": 0.0,
+      "step": 7850
+    },
+    {
+      "epoch": 0.05171052631578947,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 1.457957967122396,
+      "learning_rate": 0.0001,
+      "loss": 3.3632,
+      "loss/crossentropy": 2.236839824914932,
+      "loss/hidden": 2.946875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28002799302339554,
+      "loss/reg": 0.0,
+      "step": 7860
+    },
+    {
+      "epoch": 0.051776315789473684,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 7.502855428059896,
+      "learning_rate": 0.0001,
+      "loss": 3.3749,
+      "loss/crossentropy": 2.4405242681503294,
+      "loss/hidden": 2.9625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2591508060693741,
+      "loss/reg": 0.0,
+      "step": 7870
+    },
+    {
+      "epoch": 0.0518421052631579,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 3.8793904622395834,
+      "learning_rate": 0.0001,
+      "loss": 3.3791,
+      "loss/crossentropy": 2.5814385414123535,
+      "loss/hidden": 2.953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31668367683887483,
+      "loss/reg": 0.0,
+      "step": 7880
+    },
+    {
+      "epoch": 0.051907894736842104,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 3.8120432535807294,
+      "learning_rate": 0.0001,
+      "loss": 3.4478,
+      "loss/crossentropy": 2.185117280483246,
+      "loss/hidden": 3.234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34759806394577025,
+      "loss/reg": 0.0,
+      "step": 7890
+    },
+    {
+      "epoch": 0.05197368421052632,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.09302469889322916,
+      "learning_rate": 0.0001,
+      "loss": 3.3764,
+      "loss/crossentropy": 2.249006187915802,
+      "loss/hidden": 3.253125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28100676983594897,
+      "loss/reg": 0.0,
+      "step": 7900
+    },
+    {
+      "epoch": 0.052039473684210524,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.08615620930989583,
+      "learning_rate": 0.0001,
+      "loss": 3.261,
+      "loss/crossentropy": 2.297496807575226,
+      "loss/hidden": 3.0953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30758740454912187,
+      "loss/reg": 0.0,
+      "step": 7910
+    },
+    {
+      "epoch": 0.05210526315789474,
+      "grad_norm": 3.46875,
+      "grad_norm_var": 0.08404947916666666,
+      "learning_rate": 0.0001,
+      "loss": 3.3132,
+      "loss/crossentropy": 2.411357748508453,
+      "loss/hidden": 3.0078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29194266349077225,
+      "loss/reg": 0.0,
+      "step": 7920
+    },
+    {
+      "epoch": 0.052171052631578944,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.27255859375,
+      "learning_rate": 0.0001,
+      "loss": 3.3767,
+      "loss/crossentropy": 2.6010719895362855,
+      "loss/hidden": 2.9671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28802800476551055,
+      "loss/reg": 0.0,
+      "step": 7930
+    },
+    {
+      "epoch": 0.05223684210526316,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.0465484619140625,
+      "learning_rate": 0.0001,
+      "loss": 3.3193,
+      "loss/crossentropy": 2.6153851985931396,
+      "loss/hidden": 3.128125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2830176457762718,
+      "loss/reg": 0.0,
+      "step": 7940
+    },
+    {
+      "epoch": 0.05230263157894737,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.0275054931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.2535,
+      "loss/crossentropy": 2.2821019262075426,
+      "loss/hidden": 3.0609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28540263772010804,
+      "loss/reg": 0.0,
+      "step": 7950
+    },
+    {
+      "epoch": 0.05236842105263158,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.16284077962239582,
+      "learning_rate": 0.0001,
+      "loss": 3.2614,
+      "loss/crossentropy": 2.1502284169197083,
+      "loss/hidden": 3.134375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.305269892513752,
+      "loss/reg": 0.0,
+      "step": 7960
+    },
+    {
+      "epoch": 0.05243421052631579,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.1658843994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.3904,
+      "loss/crossentropy": 2.4389883518218993,
+      "loss/hidden": 3.1078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3048334762454033,
+      "loss/reg": 0.0,
+      "step": 7970
+    },
+    {
+      "epoch": 0.0525,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.07778218587239584,
+      "learning_rate": 0.0001,
+      "loss": 3.4152,
+      "loss/crossentropy": 2.3947508692741395,
+      "loss/hidden": 3.15625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3298331335186958,
+      "loss/reg": 0.0,
+      "step": 7980
+    },
+    {
+      "epoch": 0.05256578947368421,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.08662109375,
+      "learning_rate": 0.0001,
+      "loss": 3.3709,
+      "loss/crossentropy": 2.2743655920028685,
+      "loss/hidden": 3.0734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3255280390381813,
+      "loss/reg": 0.0,
+      "step": 7990
+    },
+    {
+      "epoch": 0.05263157894736842,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.036881510416666666,
+      "learning_rate": 0.0001,
+      "loss": 3.251,
+      "loss/crossentropy": 2.4580028295516967,
+      "loss/hidden": 2.978125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29261877238750456,
+      "loss/reg": 0.0,
+      "step": 8000
+    },
+    {
+      "epoch": 0.05269736842105263,
+      "grad_norm": 3.28125,
+      "grad_norm_var": 0.23829752604166668,
+      "learning_rate": 0.0001,
+      "loss": 3.3925,
+      "loss/crossentropy": 2.569744038581848,
+      "loss/hidden": 3.3046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4232485115528107,
+      "loss/reg": 0.0,
+      "step": 8010
+    },
+    {
+      "epoch": 0.052763157894736845,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.09871317545572916,
+      "learning_rate": 0.0001,
+      "loss": 3.33,
+      "loss/crossentropy": 2.381676936149597,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2768414840102196,
+      "loss/reg": 0.0,
+      "step": 8020
+    },
+    {
+      "epoch": 0.05282894736842105,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.07251688639322916,
+      "learning_rate": 0.0001,
+      "loss": 3.3659,
+      "loss/crossentropy": 2.529127871990204,
+      "loss/hidden": 3.046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3230110973119736,
+      "loss/reg": 0.0,
+      "step": 8030
+    },
+    {
+      "epoch": 0.052894736842105265,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.03585611979166667,
+      "learning_rate": 0.0001,
+      "loss": 3.241,
+      "loss/crossentropy": 2.4573261976242065,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26054717898368834,
+      "loss/reg": 0.0,
+      "step": 8040
+    },
+    {
+      "epoch": 0.05296052631578947,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.028319295247395834,
+      "learning_rate": 0.0001,
+      "loss": 3.3654,
+      "loss/crossentropy": 2.3388527154922487,
+      "loss/hidden": 3.03125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3071700781583786,
+      "loss/reg": 0.0,
+      "step": 8050
+    },
+    {
+      "epoch": 0.053026315789473685,
+      "grad_norm": 3.015625,
+      "grad_norm_var": 0.0393218994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.3602,
+      "loss/crossentropy": 2.4984395027160646,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2504978612065315,
+      "loss/reg": 0.0,
+      "step": 8060
+    },
+    {
+      "epoch": 0.05309210526315789,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 2.4539998372395835,
+      "learning_rate": 0.0001,
+      "loss": 3.4446,
+      "loss/crossentropy": 2.3318467855453493,
+      "loss/hidden": 3.25625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38549562990665437,
+      "loss/reg": 0.0,
+      "step": 8070
+    },
+    {
+      "epoch": 0.053157894736842105,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 2.5296946207682294,
+      "learning_rate": 0.0001,
+      "loss": 3.3183,
+      "loss/crossentropy": 2.3632636427879334,
+      "loss/hidden": 2.98125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28499974459409716,
+      "loss/reg": 0.0,
+      "step": 8080
+    },
+    {
+      "epoch": 0.05322368421052632,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.2263824462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.3982,
+      "loss/crossentropy": 2.5644856214523317,
+      "loss/hidden": 3.0421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32271777242422106,
+      "loss/reg": 0.0,
+      "step": 8090
+    },
+    {
+      "epoch": 0.053289473684210525,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.20636393229166666,
+      "learning_rate": 0.0001,
+      "loss": 3.3351,
+      "loss/crossentropy": 2.195914793014526,
+      "loss/hidden": 3.2546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32049285918474196,
+      "loss/reg": 0.0,
+      "step": 8100
+    },
+    {
+      "epoch": 0.05335526315789474,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.03369852701822917,
+      "learning_rate": 0.0001,
+      "loss": 3.3031,
+      "loss/crossentropy": 2.6255326747894285,
+      "loss/hidden": 2.940625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27993575036525725,
+      "loss/reg": 0.0,
+      "step": 8110
+    },
+    {
+      "epoch": 0.053421052631578946,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.07769775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.3496,
+      "loss/crossentropy": 2.430241084098816,
+      "loss/hidden": 3.0859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31150197684764863,
+      "loss/reg": 0.0,
+      "step": 8120
+    },
+    {
+      "epoch": 0.05348684210526316,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.0697174072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.3655,
+      "loss/crossentropy": 2.4684382557868956,
+      "loss/hidden": 2.9921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2934414252638817,
+      "loss/reg": 0.0,
+      "step": 8130
+    },
+    {
+      "epoch": 0.053552631578947366,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.09716389973958334,
+      "learning_rate": 0.0001,
+      "loss": 3.3756,
+      "loss/crossentropy": 2.679113733768463,
+      "loss/hidden": 3.3,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35277644991874696,
+      "loss/reg": 0.0,
+      "step": 8140
+    },
+    {
+      "epoch": 0.05361842105263158,
+      "grad_norm": 4.625,
+      "grad_norm_var": 0.3069986979166667,
+      "learning_rate": 0.0001,
+      "loss": 3.3979,
+      "loss/crossentropy": 2.3737685680389404,
+      "loss/hidden": 3.3359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4402651429176331,
+      "loss/reg": 0.0,
+      "step": 8150
+    },
+    {
+      "epoch": 0.05368421052631579,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.4447255452473958,
+      "learning_rate": 0.0001,
+      "loss": 3.3395,
+      "loss/crossentropy": 2.556090760231018,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2898376002907753,
+      "loss/reg": 0.0,
+      "step": 8160
+    },
+    {
+      "epoch": 0.05375,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.4153717041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.4755,
+      "loss/crossentropy": 2.3199184775352477,
+      "loss/hidden": 3.2375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3026577115058899,
+      "loss/reg": 0.0,
+      "step": 8170
+    },
+    {
+      "epoch": 0.05381578947368421,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.11787821451822916,
+      "learning_rate": 0.0001,
+      "loss": 3.3616,
+      "loss/crossentropy": 2.566536474227905,
+      "loss/hidden": 2.953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3062845066189766,
+      "loss/reg": 0.0,
+      "step": 8180
+    },
+    {
+      "epoch": 0.05388157894736842,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.0469146728515625,
+      "learning_rate": 0.0001,
+      "loss": 3.3019,
+      "loss/crossentropy": 2.434731423854828,
+      "loss/hidden": 2.9328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27008683085441587,
+      "loss/reg": 0.0,
+      "step": 8190
+    },
+    {
+      "epoch": 0.05394736842105263,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.2584299723307292,
+      "learning_rate": 0.0001,
+      "loss": 3.359,
+      "loss/crossentropy": 2.2641067147254943,
+      "loss/hidden": 3.0171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27225579768419267,
+      "loss/reg": 0.0,
+      "step": 8200
+    },
+    {
+      "epoch": 0.05401315789473684,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.27990620930989585,
+      "learning_rate": 0.0001,
+      "loss": 3.3185,
+      "loss/crossentropy": 2.410356640815735,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24587155133485794,
+      "loss/reg": 0.0,
+      "step": 8210
+    },
+    {
+      "epoch": 0.05407894736842105,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.03493550618489583,
+      "learning_rate": 0.0001,
+      "loss": 3.407,
+      "loss/crossentropy": 2.3557684421539307,
+      "loss/hidden": 2.9828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27189340591430666,
+      "loss/reg": 0.0,
+      "step": 8220
+    },
+    {
+      "epoch": 0.054144736842105266,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.031615193684895834,
+      "learning_rate": 0.0001,
+      "loss": 3.3054,
+      "loss/crossentropy": 2.453370213508606,
+      "loss/hidden": 3.0375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3121939614415169,
+      "loss/reg": 0.0,
+      "step": 8230
+    },
+    {
+      "epoch": 0.05421052631578947,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.06060791015625,
+      "learning_rate": 0.0001,
+      "loss": 3.3556,
+      "loss/crossentropy": 2.3469805240631105,
+      "loss/hidden": 3.115625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35667684078216555,
+      "loss/reg": 0.0,
+      "step": 8240
+    },
+    {
+      "epoch": 0.054276315789473686,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.10061442057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.3466,
+      "loss/crossentropy": 2.5584194660186768,
+      "loss/hidden": 3.0609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31178881525993346,
+      "loss/reg": 0.0,
+      "step": 8250
+    },
+    {
+      "epoch": 0.05434210526315789,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.06204020182291667,
+      "learning_rate": 0.0001,
+      "loss": 3.291,
+      "loss/crossentropy": 2.372837942838669,
+      "loss/hidden": 3.071875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27887275665998457,
+      "loss/reg": 0.0,
+      "step": 8260
+    },
+    {
+      "epoch": 0.054407894736842106,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.05986328125,
+      "learning_rate": 0.0001,
+      "loss": 3.3687,
+      "loss/crossentropy": 2.2389731287956236,
+      "loss/hidden": 3.0265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29073659181594846,
+      "loss/reg": 0.0,
+      "step": 8270
+    },
+    {
+      "epoch": 0.05447368421052631,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.0772369384765625,
+      "learning_rate": 0.0001,
+      "loss": 3.3364,
+      "loss/crossentropy": 2.130947244167328,
+      "loss/hidden": 3.140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3286518737673759,
+      "loss/reg": 0.0,
+      "step": 8280
+    },
+    {
+      "epoch": 0.05453947368421053,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.3402740478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.4311,
+      "loss/crossentropy": 2.5036970019340514,
+      "loss/hidden": 3.0765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3155761957168579,
+      "loss/reg": 0.0,
+      "step": 8290
+    },
+    {
+      "epoch": 0.05460526315789474,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.07206624348958333,
+      "learning_rate": 0.0001,
+      "loss": 3.3068,
+      "loss/crossentropy": 2.15233553647995,
+      "loss/hidden": 3.0859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2912480518221855,
+      "loss/reg": 0.0,
+      "step": 8300
+    },
+    {
+      "epoch": 0.05467105263157895,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.043290201822916666,
+      "learning_rate": 0.0001,
+      "loss": 3.3888,
+      "loss/crossentropy": 2.3679205000400545,
+      "loss/hidden": 3.2984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33265506476163864,
+      "loss/reg": 0.0,
+      "step": 8310
+    },
+    {
+      "epoch": 0.05473684210526316,
+      "grad_norm": 4.78125,
+      "grad_norm_var": 0.4022745768229167,
+      "learning_rate": 0.0001,
+      "loss": 3.4588,
+      "loss/crossentropy": 2.3284069895744324,
+      "loss/hidden": 3.15625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28553168624639513,
+      "loss/reg": 0.0,
+      "step": 8320
+    },
+    {
+      "epoch": 0.05480263157894737,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.5325480143229167,
+      "learning_rate": 0.0001,
+      "loss": 3.3491,
+      "loss/crossentropy": 2.602140688896179,
+      "loss/hidden": 2.9609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30461025387048724,
+      "loss/reg": 0.0,
+      "step": 8330
+    },
+    {
+      "epoch": 0.05486842105263158,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.21806233723958332,
+      "learning_rate": 0.0001,
+      "loss": 3.2902,
+      "loss/crossentropy": 2.3135936856269836,
+      "loss/hidden": 2.925,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24052460640668868,
+      "loss/reg": 0.0,
+      "step": 8340
+    },
+    {
+      "epoch": 0.05493421052631579,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.10383707682291667,
+      "learning_rate": 0.0001,
+      "loss": 3.4435,
+      "loss/crossentropy": 2.133499014377594,
+      "loss/hidden": 3.14375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3066937685012817,
+      "loss/reg": 0.0,
+      "step": 8350
+    },
+    {
+      "epoch": 0.055,
+      "grad_norm": 3.15625,
+      "grad_norm_var": 1.1286936442057292,
+      "learning_rate": 0.0001,
+      "loss": 3.3443,
+      "loss/crossentropy": 2.5188000440597533,
+      "loss/hidden": 3.340625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3301475077867508,
+      "loss/reg": 0.0,
+      "step": 8360
+    },
+    {
+      "epoch": 0.055065789473684214,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 1.1146230061848958,
+      "learning_rate": 0.0001,
+      "loss": 3.2979,
+      "loss/crossentropy": 2.5192033290863036,
+      "loss/hidden": 2.9296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29569360315799714,
+      "loss/reg": 0.0,
+      "step": 8370
+    },
+    {
+      "epoch": 0.05513157894736842,
+      "grad_norm": 4.28125,
+      "grad_norm_var": 0.2727691650390625,
+      "learning_rate": 0.0001,
+      "loss": 3.312,
+      "loss/crossentropy": 2.5402653098106383,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2737806707620621,
+      "loss/reg": 0.0,
+      "step": 8380
+    },
+    {
+      "epoch": 0.055197368421052634,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.9796702067057291,
+      "learning_rate": 0.0001,
+      "loss": 3.3479,
+      "loss/crossentropy": 2.264913785457611,
+      "loss/hidden": 3.0453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28024169653654096,
+      "loss/reg": 0.0,
+      "step": 8390
+    },
+    {
+      "epoch": 0.05526315789473684,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.14442952473958334,
+      "learning_rate": 0.0001,
+      "loss": 3.2227,
+      "loss/crossentropy": 2.3250611424446106,
+      "loss/hidden": 2.96875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29773685038089753,
+      "loss/reg": 0.0,
+      "step": 8400
+    },
+    {
+      "epoch": 0.055328947368421054,
+      "grad_norm": 3.453125,
+      "grad_norm_var": 0.1592681884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.3882,
+      "loss/crossentropy": 2.73007869720459,
+      "loss/hidden": 3.3,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.361674590408802,
+      "loss/reg": 0.0,
+      "step": 8410
+    },
+    {
+      "epoch": 0.05539473684210526,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.12535807291666667,
+      "learning_rate": 0.0001,
+      "loss": 3.3358,
+      "loss/crossentropy": 2.533698391914368,
+      "loss/hidden": 3.0359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32232231795787813,
+      "loss/reg": 0.0,
+      "step": 8420
+    },
+    {
+      "epoch": 0.055460526315789474,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 413.9550415039063,
+      "learning_rate": 0.0001,
+      "loss": 3.4177,
+      "loss/crossentropy": 2.4008097648620605,
+      "loss/hidden": 3.0328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31593555510044097,
+      "loss/reg": 0.0,
+      "step": 8430
+    },
+    {
+      "epoch": 0.05552631578947369,
+      "grad_norm": 3.234375,
+      "grad_norm_var": 413.1215077718099,
+      "learning_rate": 0.0001,
+      "loss": 3.3581,
+      "loss/crossentropy": 2.4226332664489747,
+      "loss/hidden": 3.1125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25367428809404374,
+      "loss/reg": 0.0,
+      "step": 8440
+    },
+    {
+      "epoch": 0.055592105263157894,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.14670817057291666,
+      "learning_rate": 0.0001,
+      "loss": 3.3078,
+      "loss/crossentropy": 2.630770039558411,
+      "loss/hidden": 2.996875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3664619579911232,
+      "loss/reg": 0.0,
+      "step": 8450
+    },
+    {
+      "epoch": 0.05565789473684211,
+      "grad_norm": 6.75,
+      "grad_norm_var": 2.27388916015625,
+      "learning_rate": 0.0001,
+      "loss": 3.503,
+      "loss/crossentropy": 2.376649534702301,
+      "loss/hidden": 3.1640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3476260006427765,
+      "loss/reg": 0.0,
+      "step": 8460
+    },
+    {
+      "epoch": 0.055723684210526314,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 2.541617838541667,
+      "learning_rate": 0.0001,
+      "loss": 3.4543,
+      "loss/crossentropy": 2.638149607181549,
+      "loss/hidden": 3.34375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34109789580106736,
+      "loss/reg": 0.0,
+      "step": 8470
+    },
+    {
+      "epoch": 0.05578947368421053,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.4665679931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.5013,
+      "loss/crossentropy": 2.40644109249115,
+      "loss/hidden": 3.1515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3092473894357681,
+      "loss/reg": 0.0,
+      "step": 8480
+    },
+    {
+      "epoch": 0.055855263157894734,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 1.1175282796223958,
+      "learning_rate": 0.0001,
+      "loss": 3.293,
+      "loss/crossentropy": 2.159450513124466,
+      "loss/hidden": 3.0125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27864499390125275,
+      "loss/reg": 0.0,
+      "step": 8490
+    },
+    {
+      "epoch": 0.05592105263157895,
+      "grad_norm": 4.90625,
+      "grad_norm_var": 0.49722900390625,
+      "learning_rate": 0.0001,
+      "loss": 3.3045,
+      "loss/crossentropy": 2.355340528488159,
+      "loss/hidden": 2.840625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2977081388235092,
+      "loss/reg": 0.0,
+      "step": 8500
+    },
+    {
+      "epoch": 0.05598684210526316,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.4117502848307292,
+      "learning_rate": 0.0001,
+      "loss": 3.3273,
+      "loss/crossentropy": 2.284228873252869,
+      "loss/hidden": 3.046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27203311026096344,
+      "loss/reg": 0.0,
+      "step": 8510
+    },
+    {
+      "epoch": 0.05605263157894737,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.15953776041666667,
+      "learning_rate": 0.0001,
+      "loss": 3.3038,
+      "loss/crossentropy": 2.6644370317459107,
+      "loss/hidden": 3.03125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2977398321032524,
+      "loss/reg": 0.0,
+      "step": 8520
+    },
+    {
+      "epoch": 0.05611842105263158,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.07891337076822917,
+      "learning_rate": 0.0001,
+      "loss": 3.3066,
+      "loss/crossentropy": 2.480556678771973,
+      "loss/hidden": 3.0,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3119770348072052,
+      "loss/reg": 0.0,
+      "step": 8530
+    },
+    {
+      "epoch": 0.05618421052631579,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.2862630208333333,
+      "learning_rate": 0.0001,
+      "loss": 3.3849,
+      "loss/crossentropy": 2.3611693739891053,
+      "loss/hidden": 2.909375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27296979874372485,
+      "loss/reg": 0.0,
+      "step": 8540
+    },
+    {
+      "epoch": 0.05625,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.41441141764322914,
+      "learning_rate": 0.0001,
+      "loss": 3.3496,
+      "loss/crossentropy": 2.3446286380290986,
+      "loss/hidden": 3.0125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28327286094427107,
+      "loss/reg": 0.0,
+      "step": 8550
+    },
+    {
+      "epoch": 0.05631578947368421,
+      "grad_norm": 2.625,
+      "grad_norm_var": 6.07724609375,
+      "learning_rate": 0.0001,
+      "loss": 3.4711,
+      "loss/crossentropy": 1.7737745344638824,
+      "loss/hidden": 3.171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27109832018613816,
+      "loss/reg": 0.0,
+      "step": 8560
+    },
+    {
+      "epoch": 0.05638157894736842,
+      "grad_norm": 2.875,
+      "grad_norm_var": 5.751741536458334,
+      "learning_rate": 0.0001,
+      "loss": 3.3501,
+      "loss/crossentropy": 2.3749794125556947,
+      "loss/hidden": 3.0578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2803412050008774,
+      "loss/reg": 0.0,
+      "step": 8570
+    },
+    {
+      "epoch": 0.056447368421052635,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.09153645833333333,
+      "learning_rate": 0.0001,
+      "loss": 3.3793,
+      "loss/crossentropy": 2.348732423782349,
+      "loss/hidden": 3.003125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2703657791018486,
+      "loss/reg": 0.0,
+      "step": 8580
+    },
+    {
+      "epoch": 0.05651315789473684,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 0.14538472493489582,
+      "learning_rate": 0.0001,
+      "loss": 3.3984,
+      "loss/crossentropy": 2.2422220349311828,
+      "loss/hidden": 3.04375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2956344410777092,
+      "loss/reg": 0.0,
+      "step": 8590
+    },
+    {
+      "epoch": 0.056578947368421055,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.07770894368489584,
+      "learning_rate": 0.0001,
+      "loss": 3.3787,
+      "loss/crossentropy": 2.3457955360412597,
+      "loss/hidden": 3.2359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3119618773460388,
+      "loss/reg": 0.0,
+      "step": 8600
+    },
+    {
+      "epoch": 0.05664473684210526,
+      "grad_norm": 3.40625,
+      "grad_norm_var": 1.7676717122395833,
+      "learning_rate": 0.0001,
+      "loss": 3.4129,
+      "loss/crossentropy": 2.3159496188163757,
+      "loss/hidden": 3.2421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3751305788755417,
+      "loss/reg": 0.0,
+      "step": 8610
+    },
+    {
+      "epoch": 0.056710526315789475,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 1.6932902018229166,
+      "learning_rate": 0.0001,
+      "loss": 3.4391,
+      "loss/crossentropy": 2.6694117546081544,
+      "loss/hidden": 3.053125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29238851368427277,
+      "loss/reg": 0.0,
+      "step": 8620
+    },
+    {
+      "epoch": 0.05677631578947368,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.11296284993489583,
+      "learning_rate": 0.0001,
+      "loss": 3.3208,
+      "loss/crossentropy": 2.1584444522857664,
+      "loss/hidden": 3.0796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2903019651770592,
+      "loss/reg": 0.0,
+      "step": 8630
+    },
+    {
+      "epoch": 0.056842105263157895,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.12463785807291666,
+      "learning_rate": 0.0001,
+      "loss": 3.3435,
+      "loss/crossentropy": 2.2420501947402953,
+      "loss/hidden": 3.0,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27747504264116285,
+      "loss/reg": 0.0,
+      "step": 8640
+    },
+    {
+      "epoch": 0.05690789473684211,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.16119384765625,
+      "learning_rate": 0.0001,
+      "loss": 3.226,
+      "loss/crossentropy": 2.378132700920105,
+      "loss/hidden": 2.871875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27239564061164856,
+      "loss/reg": 0.0,
+      "step": 8650
+    },
+    {
+      "epoch": 0.056973684210526315,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.06622721354166666,
+      "learning_rate": 0.0001,
+      "loss": 3.3322,
+      "loss/crossentropy": 2.1032593488693236,
+      "loss/hidden": 2.9859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2557524010539055,
+      "loss/reg": 0.0,
+      "step": 8660
+    },
+    {
+      "epoch": 0.05703947368421053,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 1.2355143229166667,
+      "learning_rate": 0.0001,
+      "loss": 3.3289,
+      "loss/crossentropy": 2.41143513917923,
+      "loss/hidden": 2.925,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27892655730247495,
+      "loss/reg": 0.0,
+      "step": 8670
+    },
+    {
+      "epoch": 0.057105263157894735,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.021117146809895834,
+      "learning_rate": 0.0001,
+      "loss": 3.348,
+      "loss/crossentropy": 2.443818140029907,
+      "loss/hidden": 2.9359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3347576320171356,
+      "loss/reg": 0.0,
+      "step": 8680
+    },
+    {
+      "epoch": 0.05717105263157895,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 9.24689275122101e+16,
+      "learning_rate": 0.0001,
+      "loss": 3.3877,
+      "loss/crossentropy": 2.181317722797394,
+      "loss/hidden": 2.9796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30654960721731184,
+      "loss/reg": 0.0,
+      "step": 8690
+    },
+    {
+      "epoch": 0.057236842105263155,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.108984375,
+      "learning_rate": 0.0001,
+      "loss": 3.308,
+      "loss/crossentropy": 2.2454151153564452,
+      "loss/hidden": 3.240625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3219583719968796,
+      "loss/reg": 0.0,
+      "step": 8700
+    },
+    {
+      "epoch": 0.05730263157894737,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.44682515462239586,
+      "learning_rate": 0.0001,
+      "loss": 3.2909,
+      "loss/crossentropy": 2.226694929599762,
+      "loss/hidden": 3.0453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27956253886222837,
+      "loss/reg": 0.0,
+      "step": 8710
+    },
+    {
+      "epoch": 0.057368421052631575,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.46708577473958335,
+      "learning_rate": 0.0001,
+      "loss": 3.3322,
+      "loss/crossentropy": 2.329686003923416,
+      "loss/hidden": 2.93125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27385311424732206,
+      "loss/reg": 0.0,
+      "step": 8720
+    },
+    {
+      "epoch": 0.05743421052631579,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.09485575358072916,
+      "learning_rate": 0.0001,
+      "loss": 3.3243,
+      "loss/crossentropy": 2.262730371952057,
+      "loss/hidden": 2.9453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26381057798862456,
+      "loss/reg": 0.0,
+      "step": 8730
+    },
+    {
+      "epoch": 0.0575,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.07937825520833333,
+      "learning_rate": 0.0001,
+      "loss": 3.3226,
+      "loss/crossentropy": 2.43484423160553,
+      "loss/hidden": 2.94375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2970830351114273,
+      "loss/reg": 0.0,
+      "step": 8740
+    },
+    {
+      "epoch": 0.05756578947368421,
+      "grad_norm": 4.0625,
+      "grad_norm_var": 0.23977864583333333,
+      "learning_rate": 0.0001,
+      "loss": 3.3517,
+      "loss/crossentropy": 2.2560600876808166,
+      "loss/hidden": 3.0796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2825756400823593,
+      "loss/reg": 0.0,
+      "step": 8750
+    },
+    {
+      "epoch": 0.05763157894736842,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 4.011067708333333,
+      "learning_rate": 0.0001,
+      "loss": 3.4317,
+      "loss/crossentropy": 2.1193652033805845,
+      "loss/hidden": 3.009375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2872270569205284,
+      "loss/reg": 0.0,
+      "step": 8760
+    },
+    {
+      "epoch": 0.05769736842105263,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.98638916015625,
+      "learning_rate": 0.0001,
+      "loss": 3.2657,
+      "loss/crossentropy": 2.2515121579170225,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2510280326008797,
+      "loss/reg": 0.0,
+      "step": 8770
+    },
+    {
+      "epoch": 0.05776315789473684,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 1.39010009765625,
+      "learning_rate": 0.0001,
+      "loss": 3.3069,
+      "loss/crossentropy": 2.45401873588562,
+      "loss/hidden": 2.975,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3077335625886917,
+      "loss/reg": 0.0,
+      "step": 8780
+    },
+    {
+      "epoch": 0.05782894736842105,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 1.4149373372395833,
+      "learning_rate": 0.0001,
+      "loss": 3.3864,
+      "loss/crossentropy": 2.3450352430343626,
+      "loss/hidden": 3.421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3485978364944458,
+      "loss/reg": 0.0,
+      "step": 8790
+    },
+    {
+      "epoch": 0.05789473684210526,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.23061421712239583,
+      "learning_rate": 0.0001,
+      "loss": 3.261,
+      "loss/crossentropy": 2.4229711413383486,
+      "loss/hidden": 3.075,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3234298795461655,
+      "loss/reg": 0.0,
+      "step": 8800
+    },
+    {
+      "epoch": 0.057960526315789476,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 4.344071451822916,
+      "learning_rate": 0.0001,
+      "loss": 3.2778,
+      "loss/crossentropy": 2.202963078022003,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2559796661138535,
+      "loss/reg": 0.0,
+      "step": 8810
+    },
+    {
+      "epoch": 0.05802631578947368,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 4.178641764322917,
+      "learning_rate": 0.0001,
+      "loss": 3.3628,
+      "loss/crossentropy": 2.2936235070228577,
+      "loss/hidden": 3.003125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32534674406051634,
+      "loss/reg": 0.0,
+      "step": 8820
+    },
+    {
+      "epoch": 0.058092105263157896,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.17635091145833334,
+      "learning_rate": 0.0001,
+      "loss": 3.3078,
+      "loss/crossentropy": 2.535597097873688,
+      "loss/hidden": 3.1109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30167998671531676,
+      "loss/reg": 0.0,
+      "step": 8830
+    },
+    {
+      "epoch": 0.0581578947368421,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.10144856770833334,
+      "learning_rate": 0.0001,
+      "loss": 3.3119,
+      "loss/crossentropy": 2.652754557132721,
+      "loss/hidden": 3.309375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37436943501234055,
+      "loss/reg": 0.0,
+      "step": 8840
+    },
+    {
+      "epoch": 0.058223684210526316,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.79921875,
+      "learning_rate": 0.0001,
+      "loss": 3.3639,
+      "loss/crossentropy": 2.4041113376617433,
+      "loss/hidden": 3.1421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.331952853500843,
+      "loss/reg": 0.0,
+      "step": 8850
+    },
+    {
+      "epoch": 0.05828947368421052,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.8512603759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.2301,
+      "loss/crossentropy": 2.7010722875595095,
+      "loss/hidden": 2.99375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2689524292945862,
+      "loss/reg": 0.0,
+      "step": 8860
+    },
+    {
+      "epoch": 0.058355263157894736,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.22805887858072918,
+      "learning_rate": 0.0001,
+      "loss": 3.301,
+      "loss/crossentropy": 2.3007858633995055,
+      "loss/hidden": 3.2953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38490410447120665,
+      "loss/reg": 0.0,
+      "step": 8870
+    },
+    {
+      "epoch": 0.05842105263157895,
+      "grad_norm": 4.375,
+      "grad_norm_var": 0.2703928629557292,
+      "learning_rate": 0.0001,
+      "loss": 3.3077,
+      "loss/crossentropy": 2.562314450740814,
+      "loss/hidden": 2.9796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29355555921792986,
+      "loss/reg": 0.0,
+      "step": 8880
+    },
+    {
+      "epoch": 0.058486842105263157,
+      "grad_norm": 4.46875,
+      "grad_norm_var": 0.5262522379557292,
+      "learning_rate": 0.0001,
+      "loss": 3.2847,
+      "loss/crossentropy": 2.329223835468292,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2784240961074829,
+      "loss/reg": 0.0,
+      "step": 8890
+    },
+    {
+      "epoch": 0.05855263157894737,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.32392476399739584,
+      "learning_rate": 0.0001,
+      "loss": 3.2975,
+      "loss/crossentropy": 2.371010947227478,
+      "loss/hidden": 3.0265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3012363612651825,
+      "loss/reg": 0.0,
+      "step": 8900
+    },
+    {
+      "epoch": 0.05861842105263158,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.05488993326822917,
+      "learning_rate": 0.0001,
+      "loss": 3.3558,
+      "loss/crossentropy": 2.4505101799964906,
+      "loss/hidden": 2.946875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28698593378067017,
+      "loss/reg": 0.0,
+      "step": 8910
+    },
+    {
+      "epoch": 0.05868421052631579,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 1.541039021809896,
+      "learning_rate": 0.0001,
+      "loss": 3.4362,
+      "loss/crossentropy": 2.569356393814087,
+      "loss/hidden": 3.009375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3198714107275009,
+      "loss/reg": 0.0,
+      "step": 8920
+    },
+    {
+      "epoch": 0.05875,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.24500325520833333,
+      "learning_rate": 0.0001,
+      "loss": 3.3947,
+      "loss/crossentropy": 2.106434017419815,
+      "loss/hidden": 2.978125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2644048437476158,
+      "loss/reg": 0.0,
+      "step": 8930
+    },
+    {
+      "epoch": 0.05881578947368421,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.21669514973958334,
+      "learning_rate": 0.0001,
+      "loss": 3.2788,
+      "loss/crossentropy": 2.517351245880127,
+      "loss/hidden": 2.9921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27984755039215087,
+      "loss/reg": 0.0,
+      "step": 8940
+    },
+    {
+      "epoch": 0.058881578947368424,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.06357421875,
+      "learning_rate": 0.0001,
+      "loss": 3.3106,
+      "loss/crossentropy": 2.3258296728134153,
+      "loss/hidden": 3.165625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32734392732381823,
+      "loss/reg": 0.0,
+      "step": 8950
+    },
+    {
+      "epoch": 0.05894736842105263,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.7411610921223958,
+      "learning_rate": 0.0001,
+      "loss": 3.4426,
+      "loss/crossentropy": 2.4165605783462523,
+      "loss/hidden": 3.125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3073220491409302,
+      "loss/reg": 0.0,
+      "step": 8960
+    },
+    {
+      "epoch": 0.059013157894736844,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.44649149576822916,
+      "learning_rate": 0.0001,
+      "loss": 3.2202,
+      "loss/crossentropy": 2.4866004467010496,
+      "loss/hidden": 2.984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30207364708185197,
+      "loss/reg": 0.0,
+      "step": 8970
+    },
+    {
+      "epoch": 0.05907894736842105,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.0566070556640625,
+      "learning_rate": 0.0001,
+      "loss": 3.2957,
+      "loss/crossentropy": 2.4917370676994324,
+      "loss/hidden": 3.078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3163344025611877,
+      "loss/reg": 0.0,
+      "step": 8980
+    },
+    {
+      "epoch": 0.059144736842105264,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.0467193603515625,
+      "learning_rate": 0.0001,
+      "loss": 3.337,
+      "loss/crossentropy": 2.529834246635437,
+      "loss/hidden": 2.90625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2896230161190033,
+      "loss/reg": 0.0,
+      "step": 8990
+    },
+    {
+      "epoch": 0.05921052631578947,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.024637858072916668,
+      "learning_rate": 0.0001,
+      "loss": 3.2664,
+      "loss/crossentropy": 2.548252558708191,
+      "loss/hidden": 3.1421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3721586674451828,
+      "loss/reg": 0.0,
+      "step": 9000
+    },
+    {
+      "epoch": 0.059276315789473684,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.010172526041666666,
+      "learning_rate": 0.0001,
+      "loss": 3.3136,
+      "loss/crossentropy": 2.628942942619324,
+      "loss/hidden": 2.86875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29181756228208544,
+      "loss/reg": 0.0,
+      "step": 9010
+    },
+    {
+      "epoch": 0.0593421052631579,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.039937337239583336,
+      "learning_rate": 0.0001,
+      "loss": 3.3276,
+      "loss/crossentropy": 2.35152667760849,
+      "loss/hidden": 2.965625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2875461965799332,
+      "loss/reg": 0.0,
+      "step": 9020
+    },
+    {
+      "epoch": 0.059407894736842104,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.0728668212890625,
+      "learning_rate": 0.0001,
+      "loss": 3.3454,
+      "loss/crossentropy": 2.3906983017921446,
+      "loss/hidden": 2.925,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2821692392230034,
+      "loss/reg": 0.0,
+      "step": 9030
+    },
+    {
+      "epoch": 0.05947368421052632,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.0730133056640625,
+      "learning_rate": 0.0001,
+      "loss": 3.3058,
+      "loss/crossentropy": 2.4100876331329344,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2886385262012482,
+      "loss/reg": 0.0,
+      "step": 9040
+    },
+    {
+      "epoch": 0.059539473684210524,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.4940592447916667,
+      "learning_rate": 0.0001,
+      "loss": 3.4398,
+      "loss/crossentropy": 2.464267885684967,
+      "loss/hidden": 3.084375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31053231209516524,
+      "loss/reg": 0.0,
+      "step": 9050
+    },
+    {
+      "epoch": 0.05960526315789474,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.08632405598958333,
+      "learning_rate": 0.0001,
+      "loss": 3.3778,
+      "loss/crossentropy": 2.402372860908508,
+      "loss/hidden": 3.0296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35591588020324705,
+      "loss/reg": 0.0,
+      "step": 9060
+    },
+    {
+      "epoch": 0.059671052631578944,
+      "grad_norm": 3.296875,
+      "grad_norm_var": 0.12502848307291667,
+      "learning_rate": 0.0001,
+      "loss": 3.3312,
+      "loss/crossentropy": 2.3369374930858613,
+      "loss/hidden": 2.871875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2521222412586212,
+      "loss/reg": 0.0,
+      "step": 9070
+    },
+    {
+      "epoch": 0.05973684210526316,
+      "grad_norm": 3.15625,
+      "grad_norm_var": 0.11041259765625,
+      "learning_rate": 0.0001,
+      "loss": 3.2977,
+      "loss/crossentropy": 2.2894081354141234,
+      "loss/hidden": 3.3359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37748522460460665,
+      "loss/reg": 0.0,
+      "step": 9080
+    },
+    {
+      "epoch": 0.05980263157894737,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 2.4495359778774493e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.4555,
+      "loss/crossentropy": 2.248319935798645,
+      "loss/hidden": 3.0234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27714093402028084,
+      "loss/reg": 0.0,
+      "step": 9090
+    },
+    {
+      "epoch": 0.05986842105263158,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 2.449535978075936e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2988,
+      "loss/crossentropy": 2.304659366607666,
+      "loss/hidden": 3.234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2838084354996681,
+      "loss/reg": 0.0,
+      "step": 9100
+    },
+    {
+      "epoch": 0.05993421052631579,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.5627838134765625,
+      "learning_rate": 0.0001,
+      "loss": 3.2864,
+      "loss/crossentropy": 2.5072904348373415,
+      "loss/hidden": 2.9890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30276854485273363,
+      "loss/reg": 0.0,
+      "step": 9110
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.5690582275390625,
+      "learning_rate": 0.0001,
+      "loss": 3.2102,
+      "loss/crossentropy": 2.2461979389190674,
+      "loss/hidden": 2.821875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2540374085307121,
+      "loss/reg": 0.0,
+      "step": 9120
+    },
+    {
+      "epoch": 0.06006578947368421,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.11433817545572916,
+      "learning_rate": 0.0001,
+      "loss": 3.2304,
+      "loss/crossentropy": 2.3598265290260314,
+      "loss/hidden": 3.0265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28421255201101303,
+      "loss/reg": 0.0,
+      "step": 9130
+    },
+    {
+      "epoch": 0.06013157894736842,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.05432535807291667,
+      "learning_rate": 0.0001,
+      "loss": 3.2493,
+      "loss/crossentropy": 2.3720561623573304,
+      "loss/hidden": 3.021875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31470999121665955,
+      "loss/reg": 0.0,
+      "step": 9140
+    },
+    {
+      "epoch": 0.06019736842105263,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.08810933430989583,
+      "learning_rate": 0.0001,
+      "loss": 3.3173,
+      "loss/crossentropy": 2.2540945589542387,
+      "loss/hidden": 2.815625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24728769958019256,
+      "loss/reg": 0.0,
+      "step": 9150
+    },
+    {
+      "epoch": 0.060263157894736845,
+      "grad_norm": 3.109375,
+      "grad_norm_var": 0.07870992024739583,
+      "learning_rate": 0.0001,
+      "loss": 3.295,
+      "loss/crossentropy": 2.2703887224197388,
+      "loss/hidden": 3.20625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27733070850372316,
+      "loss/reg": 0.0,
+      "step": 9160
+    },
+    {
+      "epoch": 0.06032894736842105,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.20073954264322916,
+      "learning_rate": 0.0001,
+      "loss": 3.2337,
+      "loss/crossentropy": 2.5513323664665224,
+      "loss/hidden": 2.8859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27180006355047226,
+      "loss/reg": 0.0,
+      "step": 9170
+    },
+    {
+      "epoch": 0.060394736842105265,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.6225545247395833,
+      "learning_rate": 0.0001,
+      "loss": 3.4001,
+      "loss/crossentropy": 2.1638057589530946,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2586470812559128,
+      "loss/reg": 0.0,
+      "step": 9180
+    },
+    {
+      "epoch": 0.06046052631578947,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.37139867146809896,
+      "learning_rate": 0.0001,
+      "loss": 3.2305,
+      "loss/crossentropy": 2.4451101064682006,
+      "loss/hidden": 2.95625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3055331766605377,
+      "loss/reg": 0.0,
+      "step": 9190
+    },
+    {
+      "epoch": 0.060526315789473685,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.32624282836914065,
+      "learning_rate": 0.0001,
+      "loss": 3.2254,
+      "loss/crossentropy": 2.561469316482544,
+      "loss/hidden": 2.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2592929035425186,
+      "loss/reg": 0.0,
+      "step": 9200
+    },
+    {
+      "epoch": 0.06059210526315789,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 1.0789459228515625,
+      "learning_rate": 0.0001,
+      "loss": 3.3551,
+      "loss/crossentropy": 2.1200980842113495,
+      "loss/hidden": 2.965625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2634043380618095,
+      "loss/reg": 0.0,
+      "step": 9210
+    },
+    {
+      "epoch": 0.060657894736842105,
+      "grad_norm": 2.0,
+      "grad_norm_var": 0.54371337890625,
+      "learning_rate": 0.0001,
+      "loss": 3.3742,
+      "loss/crossentropy": 2.4880159854888917,
+      "loss/hidden": 3.140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29063448309898376,
+      "loss/reg": 0.0,
+      "step": 9220
+    },
+    {
+      "epoch": 0.06072368421052632,
+      "grad_norm": 2.953125,
+      "grad_norm_var": 0.24761962890625,
+      "learning_rate": 0.0001,
+      "loss": 3.3781,
+      "loss/crossentropy": 2.558793306350708,
+      "loss/hidden": 3.040625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3282888367772102,
+      "loss/reg": 0.0,
+      "step": 9230
+    },
+    {
+      "epoch": 0.060789473684210525,
+      "grad_norm": 4.75,
+      "grad_norm_var": 0.35944010416666666,
+      "learning_rate": 0.0001,
+      "loss": 3.2908,
+      "loss/crossentropy": 2.3837397813797,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29975869655609133,
+      "loss/reg": 0.0,
+      "step": 9240
+    },
+    {
+      "epoch": 0.06085526315789474,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.41629130045572915,
+      "learning_rate": 0.0001,
+      "loss": 3.3291,
+      "loss/crossentropy": 2.395397412776947,
+      "loss/hidden": 3.06875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3036083221435547,
+      "loss/reg": 0.0,
+      "step": 9250
+    },
+    {
+      "epoch": 0.060921052631578945,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.09286702473958333,
+      "learning_rate": 0.0001,
+      "loss": 3.2991,
+      "loss/crossentropy": 2.350081342458725,
+      "loss/hidden": 3.234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31170106381177903,
+      "loss/reg": 0.0,
+      "step": 9260
+    },
+    {
+      "epoch": 0.06098684210526316,
+      "grad_norm": 3.125,
+      "grad_norm_var": 0.09381510416666666,
+      "learning_rate": 0.0001,
+      "loss": 3.2234,
+      "loss/crossentropy": 2.023473250865936,
+      "loss/hidden": 3.01875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2770851716399193,
+      "loss/reg": 0.0,
+      "step": 9270
+    },
+    {
+      "epoch": 0.061052631578947365,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.09631245930989583,
+      "learning_rate": 0.0001,
+      "loss": 3.2279,
+      "loss/crossentropy": 2.362475335597992,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24814699292182923,
+      "loss/reg": 0.0,
+      "step": 9280
+    },
+    {
+      "epoch": 0.06111842105263158,
+      "grad_norm": 3.1875,
+      "grad_norm_var": 0.49576416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.4035,
+      "loss/crossentropy": 2.5755128145217894,
+      "loss/hidden": 3.1328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35746499747037885,
+      "loss/reg": 0.0,
+      "step": 9290
+    },
+    {
+      "epoch": 0.06118421052631579,
+      "grad_norm": 2.953125,
+      "grad_norm_var": 0.09010009765625,
+      "learning_rate": 0.0001,
+      "loss": 3.2592,
+      "loss/crossentropy": 2.2949343085289002,
+      "loss/hidden": 3.08125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2996257975697517,
+      "loss/reg": 0.0,
+      "step": 9300
+    },
+    {
+      "epoch": 0.06125,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.3297190348307292,
+      "learning_rate": 0.0001,
+      "loss": 3.3329,
+      "loss/crossentropy": 2.4506813049316407,
+      "loss/hidden": 3.09375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3265557274222374,
+      "loss/reg": 0.0,
+      "step": 9310
+    },
+    {
+      "epoch": 0.06131578947368421,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.3309529622395833,
+      "learning_rate": 0.0001,
+      "loss": 3.2163,
+      "loss/crossentropy": 2.3078281760215758,
+      "loss/hidden": 3.0609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2651080995798111,
+      "loss/reg": 0.0,
+      "step": 9320
+    },
+    {
+      "epoch": 0.06138157894736842,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.26513671875,
+      "learning_rate": 0.0001,
+      "loss": 3.2706,
+      "loss/crossentropy": 2.1937548160552978,
+      "loss/hidden": 2.98125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2700933560729027,
+      "loss/reg": 0.0,
+      "step": 9330
+    },
+    {
+      "epoch": 0.06144736842105263,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.31164957682291666,
+      "learning_rate": 0.0001,
+      "loss": 3.4252,
+      "loss/crossentropy": 2.112582105398178,
+      "loss/hidden": 3.0984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26334969997406005,
+      "loss/reg": 0.0,
+      "step": 9340
+    },
+    {
+      "epoch": 0.06151315789473684,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.33046875,
+      "learning_rate": 0.0001,
+      "loss": 3.3277,
+      "loss/crossentropy": 2.4319912672042845,
+      "loss/hidden": 2.9765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2982776537537575,
+      "loss/reg": 0.0,
+      "step": 9350
+    },
+    {
+      "epoch": 0.06157894736842105,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.48318684895833336,
+      "learning_rate": 0.0001,
+      "loss": 3.28,
+      "loss/crossentropy": 2.3863817691802978,
+      "loss/hidden": 2.9515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2783980667591095,
+      "loss/reg": 0.0,
+      "step": 9360
+    },
+    {
+      "epoch": 0.061644736842105266,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.4833730061848958,
+      "learning_rate": 0.0001,
+      "loss": 3.329,
+      "loss/crossentropy": 2.5484490633010863,
+      "loss/hidden": 3.03125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3031032904982567,
+      "loss/reg": 0.0,
+      "step": 9370
+    },
+    {
+      "epoch": 0.06171052631578947,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.08167317708333334,
+      "learning_rate": 0.0001,
+      "loss": 3.2587,
+      "loss/crossentropy": 2.2674924612045286,
+      "loss/hidden": 3.1234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3650638833642006,
+      "loss/reg": 0.0,
+      "step": 9380
+    },
+    {
+      "epoch": 0.061776315789473686,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.22155659993489582,
+      "learning_rate": 0.0001,
+      "loss": 3.3241,
+      "loss/crossentropy": 2.3348896741867065,
+      "loss/hidden": 3.1203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.283182792365551,
+      "loss/reg": 0.0,
+      "step": 9390
+    },
+    {
+      "epoch": 0.06184210526315789,
+      "grad_norm": 3.15625,
+      "grad_norm_var": 1.1829427083333333,
+      "learning_rate": 0.0001,
+      "loss": 3.3629,
+      "loss/crossentropy": 2.464947986602783,
+      "loss/hidden": 2.971875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28233895897865297,
+      "loss/reg": 0.0,
+      "step": 9400
+    },
+    {
+      "epoch": 0.061907894736842106,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 1.1680623372395833,
+      "learning_rate": 0.0001,
+      "loss": 3.3087,
+      "loss/crossentropy": 2.179221343994141,
+      "loss/hidden": 3.04375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3310952290892601,
+      "loss/reg": 0.0,
+      "step": 9410
+    },
+    {
+      "epoch": 0.06197368421052631,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 2.0582183837890624,
+      "learning_rate": 0.0001,
+      "loss": 3.4293,
+      "loss/crossentropy": 2.413185155391693,
+      "loss/hidden": 3.18125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3003757044672966,
+      "loss/reg": 0.0,
+      "step": 9420
+    },
+    {
+      "epoch": 0.062039473684210526,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.3502349853515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1251,
+      "loss/crossentropy": 2.3518580555915833,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2455562546849251,
+      "loss/reg": 0.0,
+      "step": 9430
+    },
+    {
+      "epoch": 0.06210526315789474,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.028425089518229165,
+      "learning_rate": 0.0001,
+      "loss": 3.3233,
+      "loss/crossentropy": 2.375118088722229,
+      "loss/hidden": 3.153125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31473297625780106,
+      "loss/reg": 0.0,
+      "step": 9440
+    },
+    {
+      "epoch": 0.062171052631578946,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.0333984375,
+      "learning_rate": 0.0001,
+      "loss": 3.2654,
+      "loss/crossentropy": 2.3164775133132935,
+      "loss/hidden": 2.9890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2699767738580704,
+      "loss/reg": 0.0,
+      "step": 9450
+    },
+    {
+      "epoch": 0.06223684210526316,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.0964263916015625,
+      "learning_rate": 0.0001,
+      "loss": 3.3552,
+      "loss/crossentropy": 2.233529049158096,
+      "loss/hidden": 3.21875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3262931898236275,
+      "loss/reg": 0.0,
+      "step": 9460
+    },
+    {
+      "epoch": 0.062302631578947366,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 24.440249633789062,
+      "learning_rate": 0.0001,
+      "loss": 3.2359,
+      "loss/crossentropy": 2.3248747825622558,
+      "loss/hidden": 2.8234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24342142790555954,
+      "loss/reg": 0.0,
+      "step": 9470
+    },
+    {
+      "epoch": 0.06236842105263158,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.06562398274739584,
+      "learning_rate": 0.0001,
+      "loss": 3.3237,
+      "loss/crossentropy": 2.588094711303711,
+      "loss/hidden": 3.128125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3172804355621338,
+      "loss/reg": 0.0,
+      "step": 9480
+    },
+    {
+      "epoch": 0.062434210526315786,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.21744384765625,
+      "learning_rate": 0.0001,
+      "loss": 3.3475,
+      "loss/crossentropy": 2.390676462650299,
+      "loss/hidden": 2.9640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27351657301187515,
+      "loss/reg": 0.0,
+      "step": 9490
+    },
+    {
+      "epoch": 0.0625,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.0399078369140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2819,
+      "loss/crossentropy": 2.3253311276435853,
+      "loss/hidden": 2.89375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.265305657684803,
+      "loss/reg": 0.0,
+      "step": 9500
+    },
+    {
+      "epoch": 0.06256578947368421,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.08692118326822916,
+      "learning_rate": 0.0001,
+      "loss": 3.2535,
+      "loss/crossentropy": 2.3912763714790346,
+      "loss/hidden": 2.9734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2748603358864784,
+      "loss/reg": 0.0,
+      "step": 9510
+    },
+    {
+      "epoch": 0.06263157894736843,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.11210098266601562,
+      "learning_rate": 0.0001,
+      "loss": 3.2609,
+      "loss/crossentropy": 2.3854790568351745,
+      "loss/hidden": 2.9046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2736155718564987,
+      "loss/reg": 0.0,
+      "step": 9520
+    },
+    {
+      "epoch": 0.06269736842105263,
+      "grad_norm": 2.9375,
+      "grad_norm_var": 0.13178888956705728,
+      "learning_rate": 0.0001,
+      "loss": 3.2434,
+      "loss/crossentropy": 2.4524547696113586,
+      "loss/hidden": 2.809375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.256032706797123,
+      "loss/reg": 0.0,
+      "step": 9530
+    },
+    {
+      "epoch": 0.06276315789473684,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.08687744140625,
+      "learning_rate": 0.0001,
+      "loss": 3.3256,
+      "loss/crossentropy": 2.563627076148987,
+      "loss/hidden": 3.3046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3692674309015274,
+      "loss/reg": 0.0,
+      "step": 9540
+    },
+    {
+      "epoch": 0.06282894736842105,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.13664957682291667,
+      "learning_rate": 0.0001,
+      "loss": 3.3197,
+      "loss/crossentropy": 2.3016056180000306,
+      "loss/hidden": 3.025,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2520491242408752,
+      "loss/reg": 0.0,
+      "step": 9550
+    },
+    {
+      "epoch": 0.06289473684210527,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.4147857666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.511,
+      "loss/crossentropy": 2.1577144265174866,
+      "loss/hidden": 3.1046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2929541230201721,
+      "loss/reg": 0.0,
+      "step": 9560
+    },
+    {
+      "epoch": 0.06296052631578947,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.18341471354166666,
+      "learning_rate": 0.0001,
+      "loss": 3.2549,
+      "loss/crossentropy": 2.31901068687439,
+      "loss/hidden": 2.90625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27463280111551286,
+      "loss/reg": 0.0,
+      "step": 9570
+    },
+    {
+      "epoch": 0.06302631578947368,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.4622884114583333,
+      "learning_rate": 0.0001,
+      "loss": 3.3477,
+      "loss/crossentropy": 2.6131432056427,
+      "loss/hidden": 3.1140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.38214774429798126,
+      "loss/reg": 0.0,
+      "step": 9580
+    },
+    {
+      "epoch": 0.0630921052631579,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.13547337849934896,
+      "learning_rate": 0.0001,
+      "loss": 3.2885,
+      "loss/crossentropy": 2.451016199588776,
+      "loss/hidden": 2.9875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2813758164644241,
+      "loss/reg": 0.0,
+      "step": 9590
+    },
+    {
+      "epoch": 0.06315789473684211,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.2512003580729167,
+      "learning_rate": 0.0001,
+      "loss": 3.3345,
+      "loss/crossentropy": 2.1964026927947997,
+      "loss/hidden": 3.2765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31094489246606827,
+      "loss/reg": 0.0,
+      "step": 9600
+    },
+    {
+      "epoch": 0.06322368421052632,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.28735249837239585,
+      "learning_rate": 0.0001,
+      "loss": 3.2956,
+      "loss/crossentropy": 2.3820174098014832,
+      "loss/hidden": 3.1375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30036603659391403,
+      "loss/reg": 0.0,
+      "step": 9610
+    },
+    {
+      "epoch": 0.06328947368421052,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.9716102600097656,
+      "learning_rate": 0.0001,
+      "loss": 3.2363,
+      "loss/crossentropy": 2.4219281673431396,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2639324277639389,
+      "loss/reg": 0.0,
+      "step": 9620
+    },
+    {
+      "epoch": 0.06335526315789473,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.8159016927083333,
+      "learning_rate": 0.0001,
+      "loss": 3.331,
+      "loss/crossentropy": 1.9574783891439438,
+      "loss/hidden": 3.1921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2543580986559391,
+      "loss/reg": 0.0,
+      "step": 9630
+    },
+    {
+      "epoch": 0.06342105263157895,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.06944961547851562,
+      "learning_rate": 0.0001,
+      "loss": 3.2008,
+      "loss/crossentropy": 2.210974097251892,
+      "loss/hidden": 2.9671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2627203479409218,
+      "loss/reg": 0.0,
+      "step": 9640
+    },
+    {
+      "epoch": 0.06348684210526316,
+      "grad_norm": 2.9375,
+      "grad_norm_var": 0.2388336181640625,
+      "learning_rate": 0.0001,
+      "loss": 3.3255,
+      "loss/crossentropy": 2.225372338294983,
+      "loss/hidden": 3.0984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28057117611169813,
+      "loss/reg": 0.0,
+      "step": 9650
+    },
+    {
+      "epoch": 0.06355263157894737,
+      "grad_norm": 2.0,
+      "grad_norm_var": 0.3254954020182292,
+      "learning_rate": 0.0001,
+      "loss": 3.3021,
+      "loss/crossentropy": 2.190217161178589,
+      "loss/hidden": 3.1734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31579277813434603,
+      "loss/reg": 0.0,
+      "step": 9660
+    },
+    {
+      "epoch": 0.06361842105263157,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.5466105143229166,
+      "learning_rate": 0.0001,
+      "loss": 3.3119,
+      "loss/crossentropy": 2.427161252498627,
+      "loss/hidden": 3.0046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.307109659910202,
+      "loss/reg": 0.0,
+      "step": 9670
+    },
+    {
+      "epoch": 0.06368421052631579,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.48213602701822916,
+      "learning_rate": 0.0001,
+      "loss": 3.2915,
+      "loss/crossentropy": 2.2590057969093325,
+      "loss/hidden": 3.0765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2674726366996765,
+      "loss/reg": 0.0,
+      "step": 9680
+    },
+    {
+      "epoch": 0.06375,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.2637858072916667,
+      "learning_rate": 0.0001,
+      "loss": 3.2742,
+      "loss/crossentropy": 2.3426333904266357,
+      "loss/hidden": 2.9328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2729641154408455,
+      "loss/reg": 0.0,
+      "step": 9690
+    },
+    {
+      "epoch": 0.06381578947368421,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.3402252197265625,
+      "learning_rate": 0.0001,
+      "loss": 3.2983,
+      "loss/crossentropy": 2.3633025169372557,
+      "loss/hidden": 2.9109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2770617350935936,
+      "loss/reg": 0.0,
+      "step": 9700
+    },
+    {
+      "epoch": 0.06388157894736841,
+      "grad_norm": 3.25,
+      "grad_norm_var": 0.23658447265625,
+      "learning_rate": 0.0001,
+      "loss": 3.2621,
+      "loss/crossentropy": 2.363292157649994,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2777522191405296,
+      "loss/reg": 0.0,
+      "step": 9710
+    },
+    {
+      "epoch": 0.06394736842105263,
+      "grad_norm": 3.5625,
+      "grad_norm_var": 0.45250244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.3017,
+      "loss/crossentropy": 2.4210567593574526,
+      "loss/hidden": 2.921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2831977978348732,
+      "loss/reg": 0.0,
+      "step": 9720
+    },
+    {
+      "epoch": 0.06401315789473684,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.2709147135416667,
+      "learning_rate": 0.0001,
+      "loss": 3.234,
+      "loss/crossentropy": 2.3109630227088926,
+      "loss/hidden": 2.9140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31227574199438096,
+      "loss/reg": 0.0,
+      "step": 9730
+    },
+    {
+      "epoch": 0.06407894736842105,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.38984273274739584,
+      "learning_rate": 0.0001,
+      "loss": 3.2414,
+      "loss/crossentropy": 2.225367599725723,
+      "loss/hidden": 3.334375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24756639897823335,
+      "loss/reg": 0.0,
+      "step": 9740
+    },
+    {
+      "epoch": 0.06414473684210527,
+      "grad_norm": 3.65625,
+      "grad_norm_var": 0.40685933430989585,
+      "learning_rate": 0.0001,
+      "loss": 3.2959,
+      "loss/crossentropy": 2.633290505409241,
+      "loss/hidden": 2.9546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27978702783584597,
+      "loss/reg": 0.0,
+      "step": 9750
+    },
+    {
+      "epoch": 0.06421052631578947,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 1.353466796875,
+      "learning_rate": 0.0001,
+      "loss": 3.2732,
+      "loss/crossentropy": 2.377350616455078,
+      "loss/hidden": 2.91875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27294492572546003,
+      "loss/reg": 0.0,
+      "step": 9760
+    },
+    {
+      "epoch": 0.06427631578947368,
+      "grad_norm": 2.375,
+      "grad_norm_var": 1.3017242431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.3766,
+      "loss/crossentropy": 2.230624866485596,
+      "loss/hidden": 2.909375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2353967770934105,
+      "loss/reg": 0.0,
+      "step": 9770
+    },
+    {
+      "epoch": 0.0643421052631579,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 30.074632771809895,
+      "learning_rate": 0.0001,
+      "loss": 3.2769,
+      "loss/crossentropy": 2.250266909599304,
+      "loss/hidden": 3.0203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2896205812692642,
+      "loss/reg": 0.0,
+      "step": 9780
+    },
+    {
+      "epoch": 0.06440789473684211,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 30.214286295572915,
+      "learning_rate": 0.0001,
+      "loss": 3.2666,
+      "loss/crossentropy": 2.495633435249329,
+      "loss/hidden": 2.8984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2458545297384262,
+      "loss/reg": 0.0,
+      "step": 9790
+    },
+    {
+      "epoch": 0.06447368421052632,
+      "grad_norm": 1.96875,
+      "grad_norm_var": 0.09082743326822916,
+      "learning_rate": 0.0001,
+      "loss": 3.2638,
+      "loss/crossentropy": 2.140651452541351,
+      "loss/hidden": 3.0890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27739599645137786,
+      "loss/reg": 0.0,
+      "step": 9800
+    },
+    {
+      "epoch": 0.06453947368421052,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 2.439875284830729,
+      "learning_rate": 0.0001,
+      "loss": 3.2928,
+      "loss/crossentropy": 2.273459422588348,
+      "loss/hidden": 2.8859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26713447719812394,
+      "loss/reg": 0.0,
+      "step": 9810
+    },
+    {
+      "epoch": 0.06460526315789474,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 2.7100901285807293,
+      "learning_rate": 0.0001,
+      "loss": 3.2839,
+      "loss/crossentropy": 2.4476850271224975,
+      "loss/hidden": 3.3984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35955790579319,
+      "loss/reg": 0.0,
+      "step": 9820
+    },
+    {
+      "epoch": 0.06467105263157895,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.46149800618489584,
+      "learning_rate": 0.0001,
+      "loss": 3.306,
+      "loss/crossentropy": 2.3544110536575316,
+      "loss/hidden": 2.9515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31140194088220596,
+      "loss/reg": 0.0,
+      "step": 9830
+    },
+    {
+      "epoch": 0.06473684210526316,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.15746968587239582,
+      "learning_rate": 0.0001,
+      "loss": 3.2209,
+      "loss/crossentropy": 2.2321391999721527,
+      "loss/hidden": 2.8953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2574504017829895,
+      "loss/reg": 0.0,
+      "step": 9840
+    },
+    {
+      "epoch": 0.06480263157894736,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.10312093098958333,
+      "learning_rate": 0.0001,
+      "loss": 3.2513,
+      "loss/crossentropy": 2.311957097053528,
+      "loss/hidden": 2.846875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2625372394919395,
+      "loss/reg": 0.0,
+      "step": 9850
+    },
+    {
+      "epoch": 0.06486842105263158,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.0565826416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.3056,
+      "loss/crossentropy": 2.331065666675568,
+      "loss/hidden": 2.9859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.277868440747261,
+      "loss/reg": 0.0,
+      "step": 9860
+    },
+    {
+      "epoch": 0.06493421052631579,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.18406575520833332,
+      "learning_rate": 0.0001,
+      "loss": 3.3417,
+      "loss/crossentropy": 2.314790654182434,
+      "loss/hidden": 3.0328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33730033934116366,
+      "loss/reg": 0.0,
+      "step": 9870
+    },
+    {
+      "epoch": 0.065,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.08243815104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.3522,
+      "loss/crossentropy": 2.489400041103363,
+      "loss/hidden": 2.9765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28822820335626603,
+      "loss/reg": 0.0,
+      "step": 9880
+    },
+    {
+      "epoch": 0.06506578947368422,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.0967437744140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2368,
+      "loss/crossentropy": 2.3749926924705504,
+      "loss/hidden": 3.175,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3226087599992752,
+      "loss/reg": 0.0,
+      "step": 9890
+    },
+    {
+      "epoch": 0.06513157894736842,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.11569722493489583,
+      "learning_rate": 0.0001,
+      "loss": 3.2881,
+      "loss/crossentropy": 2.190938687324524,
+      "loss/hidden": 3.0296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30352693498134614,
+      "loss/reg": 0.0,
+      "step": 9900
+    },
+    {
+      "epoch": 0.06519736842105263,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.09851888020833334,
+      "learning_rate": 0.0001,
+      "loss": 3.2849,
+      "loss/crossentropy": 2.412072277069092,
+      "loss/hidden": 3.0234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32761459052562714,
+      "loss/reg": 0.0,
+      "step": 9910
+    },
+    {
+      "epoch": 0.06526315789473684,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.49250895182291665,
+      "learning_rate": 0.0001,
+      "loss": 3.2984,
+      "loss/crossentropy": 2.2049274504184724,
+      "loss/hidden": 2.971875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2710462361574173,
+      "loss/reg": 0.0,
+      "step": 9920
+    },
+    {
+      "epoch": 0.06532894736842106,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.2756581624348958,
+      "learning_rate": 0.0001,
+      "loss": 3.3729,
+      "loss/crossentropy": 2.0697293996810915,
+      "loss/hidden": 3.3359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32963491380214693,
+      "loss/reg": 0.0,
+      "step": 9930
+    },
+    {
+      "epoch": 0.06539473684210527,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.18677978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.2972,
+      "loss/crossentropy": 2.3622434020042418,
+      "loss/hidden": 3.103125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31804386228322984,
+      "loss/reg": 0.0,
+      "step": 9940
+    },
+    {
+      "epoch": 0.06546052631578947,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.07353413899739583,
+      "learning_rate": 0.0001,
+      "loss": 3.3347,
+      "loss/crossentropy": 2.4146655321121218,
+      "loss/hidden": 3.309375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36195366978645327,
+      "loss/reg": 0.0,
+      "step": 9950
+    },
+    {
+      "epoch": 0.06552631578947368,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.031086222330729166,
+      "learning_rate": 0.0001,
+      "loss": 3.3045,
+      "loss/crossentropy": 2.499011588096619,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27686055898666384,
+      "loss/reg": 0.0,
+      "step": 9960
+    },
+    {
+      "epoch": 0.0655921052631579,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.36774800618489584,
+      "learning_rate": 0.0001,
+      "loss": 3.3505,
+      "loss/crossentropy": 2.2883435606956484,
+      "loss/hidden": 2.9421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.43403479307889936,
+      "loss/reg": 0.0,
+      "step": 9970
+    },
+    {
+      "epoch": 0.06565789473684211,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.39895426432291664,
+      "learning_rate": 0.0001,
+      "loss": 3.3004,
+      "loss/crossentropy": 2.328636658191681,
+      "loss/hidden": 2.9828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30042982697486875,
+      "loss/reg": 0.0,
+      "step": 9980
+    },
+    {
+      "epoch": 0.06572368421052631,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.070458984375,
+      "learning_rate": 0.0001,
+      "loss": 3.3201,
+      "loss/crossentropy": 2.1099472880363463,
+      "loss/hidden": 3.2375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2899234861135483,
+      "loss/reg": 0.0,
+      "step": 9990
+    },
+    {
+      "epoch": 0.06578947368421052,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.06652018229166666,
+      "learning_rate": 0.0001,
+      "loss": 3.3435,
+      "loss/crossentropy": 2.2847598433494567,
+      "loss/hidden": 2.9890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3000261425971985,
+      "loss/reg": 0.0,
+      "step": 10000
+    },
+    {
+      "epoch": 0.06585526315789474,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.08883463541666667,
+      "learning_rate": 0.0001,
+      "loss": 3.2377,
+      "loss/crossentropy": 2.4516909599304197,
+      "loss/hidden": 3.0265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3408001005649567,
+      "loss/reg": 0.0,
+      "step": 10010
+    },
+    {
+      "epoch": 0.06592105263157895,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 4.412495930989583,
+      "learning_rate": 0.0001,
+      "loss": 3.417,
+      "loss/crossentropy": 2.3393358111381533,
+      "loss/hidden": 3.153125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3251391679048538,
+      "loss/reg": 0.0,
+      "step": 10020
+    },
+    {
+      "epoch": 0.06598684210526316,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 4.318382771809896,
+      "learning_rate": 0.0001,
+      "loss": 3.318,
+      "loss/crossentropy": 2.4476661682128906,
+      "loss/hidden": 2.8234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25891698747873304,
+      "loss/reg": 0.0,
+      "step": 10030
+    },
+    {
+      "epoch": 0.06605263157894736,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 2.029248046875,
+      "learning_rate": 0.0001,
+      "loss": 3.3175,
+      "loss/crossentropy": 2.5090669870376585,
+      "loss/hidden": 3.0109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3005325973033905,
+      "loss/reg": 0.0,
+      "step": 10040
+    },
+    {
+      "epoch": 0.06611842105263158,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 18.96970926920573,
+      "learning_rate": 0.0001,
+      "loss": 3.3768,
+      "loss/crossentropy": 2.4451366662979126,
+      "loss/hidden": 2.925,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2800415620207787,
+      "loss/reg": 0.0,
+      "step": 10050
+    },
+    {
+      "epoch": 0.06618421052631579,
+      "grad_norm": 2.9375,
+      "grad_norm_var": 0.11100260416666667,
+      "learning_rate": 0.0001,
+      "loss": 3.3131,
+      "loss/crossentropy": 2.3123559236526487,
+      "loss/hidden": 3.040625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32703131139278413,
+      "loss/reg": 0.0,
+      "step": 10060
+    },
+    {
+      "epoch": 0.06625,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.055985514322916666,
+      "learning_rate": 0.0001,
+      "loss": 3.2837,
+      "loss/crossentropy": 2.20616455078125,
+      "loss/hidden": 3.059375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3035066843032837,
+      "loss/reg": 0.0,
+      "step": 10070
+    },
+    {
+      "epoch": 0.06631578947368422,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.025560506184895835,
+      "learning_rate": 0.0001,
+      "loss": 3.202,
+      "loss/crossentropy": 2.3889974474906923,
+      "loss/hidden": 2.9078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25791111290454866,
+      "loss/reg": 0.0,
+      "step": 10080
+    },
+    {
+      "epoch": 0.06638157894736842,
+      "grad_norm": 3.078125,
+      "grad_norm_var": 0.12398681640625,
+      "learning_rate": 0.0001,
+      "loss": 3.2662,
+      "loss/crossentropy": 2.0885241270065307,
+      "loss/hidden": 3.0421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26190474927425383,
+      "loss/reg": 0.0,
+      "step": 10090
+    },
+    {
+      "epoch": 0.06644736842105263,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.10334879557291667,
+      "learning_rate": 0.0001,
+      "loss": 3.2587,
+      "loss/crossentropy": 2.3084590315818785,
+      "loss/hidden": 3.0078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.292548905313015,
+      "loss/reg": 0.0,
+      "step": 10100
+    },
+    {
+      "epoch": 0.06651315789473684,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.11416727701822917,
+      "learning_rate": 0.0001,
+      "loss": 3.214,
+      "loss/crossentropy": 2.519231605529785,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2500781774520874,
+      "loss/reg": 0.0,
+      "step": 10110
+    },
+    {
+      "epoch": 0.06657894736842106,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.10711263020833334,
+      "learning_rate": 0.0001,
+      "loss": 3.3349,
+      "loss/crossentropy": 2.3334843158721923,
+      "loss/hidden": 3.034375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2946275919675827,
+      "loss/reg": 0.0,
+      "step": 10120
+    },
+    {
+      "epoch": 0.06664473684210526,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.3732086181640625,
+      "learning_rate": 0.0001,
+      "loss": 3.3638,
+      "loss/crossentropy": 2.4801114797592163,
+      "loss/hidden": 2.9609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29529736936092377,
+      "loss/reg": 0.0,
+      "step": 10130
+    },
+    {
+      "epoch": 0.06671052631578947,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.4193511962890625,
+      "learning_rate": 0.0001,
+      "loss": 3.3195,
+      "loss/crossentropy": 2.3818048357963564,
+      "loss/hidden": 3.0453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.299163943529129,
+      "loss/reg": 0.0,
+      "step": 10140
+    },
+    {
+      "epoch": 0.06677631578947368,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.07637430826822916,
+      "learning_rate": 0.0001,
+      "loss": 3.2907,
+      "loss/crossentropy": 2.3112216353416444,
+      "loss/hidden": 2.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24870002567768096,
+      "loss/reg": 0.0,
+      "step": 10150
+    },
+    {
+      "epoch": 0.0668421052631579,
+      "grad_norm": 3.859375,
+      "grad_norm_var": 0.16756083170572916,
+      "learning_rate": 0.0001,
+      "loss": 3.369,
+      "loss/crossentropy": 2.1794182300567626,
+      "loss/hidden": 3.0375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3170015588402748,
+      "loss/reg": 0.0,
+      "step": 10160
+    },
+    {
+      "epoch": 0.06690789473684211,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.17148030598958333,
+      "learning_rate": 0.0001,
+      "loss": 3.3343,
+      "loss/crossentropy": 2.125279116630554,
+      "loss/hidden": 3.084375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27327116429805753,
+      "loss/reg": 0.0,
+      "step": 10170
+    },
+    {
+      "epoch": 0.06697368421052631,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 1.8383626302083333,
+      "learning_rate": 0.0001,
+      "loss": 3.3548,
+      "loss/crossentropy": 2.4289526462554933,
+      "loss/hidden": 3.065625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28756752908229827,
+      "loss/reg": 0.0,
+      "step": 10180
+    },
+    {
+      "epoch": 0.06703947368421052,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.4429677327473958,
+      "learning_rate": 0.0001,
+      "loss": 3.3431,
+      "loss/crossentropy": 2.297343075275421,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2415475845336914,
+      "loss/reg": 0.0,
+      "step": 10190
+    },
+    {
+      "epoch": 0.06710526315789474,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.53648681640625,
+      "learning_rate": 0.0001,
+      "loss": 3.2412,
+      "loss/crossentropy": 2.291072869300842,
+      "loss/hidden": 3.0828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3272585093975067,
+      "loss/reg": 0.0,
+      "step": 10200
+    },
+    {
+      "epoch": 0.06717105263157895,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.03841044108072917,
+      "learning_rate": 0.0001,
+      "loss": 3.2881,
+      "loss/crossentropy": 2.3315325021743774,
+      "loss/hidden": 3.1046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2916677713394165,
+      "loss/reg": 0.0,
+      "step": 10210
+    },
+    {
+      "epoch": 0.06723684210526316,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.05279541015625,
+      "learning_rate": 0.0001,
+      "loss": 3.2758,
+      "loss/crossentropy": 2.496378016471863,
+      "loss/hidden": 3.06875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2993095234036446,
+      "loss/reg": 0.0,
+      "step": 10220
+    },
+    {
+      "epoch": 0.06730263157894736,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.030101521809895834,
+      "learning_rate": 0.0001,
+      "loss": 3.184,
+      "loss/crossentropy": 1.954445093870163,
+      "loss/hidden": 2.953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23346130400896073,
+      "loss/reg": 0.0,
+      "step": 10230
+    },
+    {
+      "epoch": 0.06736842105263158,
+      "grad_norm": 3.6875,
+      "grad_norm_var": 0.2775675455729167,
+      "learning_rate": 0.0001,
+      "loss": 3.2837,
+      "loss/crossentropy": 2.4431410312652586,
+      "loss/hidden": 3.021875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3258019149303436,
+      "loss/reg": 0.0,
+      "step": 10240
+    },
+    {
+      "epoch": 0.06743421052631579,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.33483784993489585,
+      "learning_rate": 0.0001,
+      "loss": 3.2903,
+      "loss/crossentropy": 2.0363747119903564,
+      "loss/hidden": 3.4796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3713301241397858,
+      "loss/reg": 0.0,
+      "step": 10250
+    },
+    {
+      "epoch": 0.0675,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.3346028645833333,
+      "learning_rate": 0.0001,
+      "loss": 3.2875,
+      "loss/crossentropy": 2.1991047143936155,
+      "loss/hidden": 3.103125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3233490988612175,
+      "loss/reg": 0.0,
+      "step": 10260
+    },
+    {
+      "epoch": 0.0675657894736842,
+      "grad_norm": 3.890625,
+      "grad_norm_var": 0.1937652587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.2766,
+      "loss/crossentropy": 2.334232974052429,
+      "loss/hidden": 3.15625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3195400908589363,
+      "loss/reg": 0.0,
+      "step": 10270
+    },
+    {
+      "epoch": 0.06763157894736842,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.7010162353515625,
+      "learning_rate": 0.0001,
+      "loss": 3.224,
+      "loss/crossentropy": 2.242165985703468,
+      "loss/hidden": 3.075,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27829615995287893,
+      "loss/reg": 0.0,
+      "step": 10280
+    },
+    {
+      "epoch": 0.06769736842105263,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.1137115478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.2283,
+      "loss/crossentropy": 2.4408915996551515,
+      "loss/hidden": 2.94375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2770886033773422,
+      "loss/reg": 0.0,
+      "step": 10290
+    },
+    {
+      "epoch": 0.06776315789473684,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.28396708170572915,
+      "learning_rate": 0.0001,
+      "loss": 3.2788,
+      "loss/crossentropy": 2.2959680914878846,
+      "loss/hidden": 2.8625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2606198683381081,
+      "loss/reg": 0.0,
+      "step": 10300
+    },
+    {
+      "epoch": 0.06782894736842106,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.0829498291015625,
+      "learning_rate": 0.0001,
+      "loss": 3.3487,
+      "loss/crossentropy": 2.430151104927063,
+      "loss/hidden": 2.921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25218638926744463,
+      "loss/reg": 0.0,
+      "step": 10310
+    },
+    {
+      "epoch": 0.06789473684210526,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.3111979166666667,
+      "learning_rate": 0.0001,
+      "loss": 3.3676,
+      "loss/crossentropy": 2.1297863006591795,
+      "loss/hidden": 3.415625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4299448400735855,
+      "loss/reg": 0.0,
+      "step": 10320
+    },
+    {
+      "epoch": 0.06796052631578947,
+      "grad_norm": 3.0625,
+      "grad_norm_var": 0.34820556640625,
+      "learning_rate": 0.0001,
+      "loss": 3.3013,
+      "loss/crossentropy": 2.2383357286453247,
+      "loss/hidden": 2.990625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31647931337356566,
+      "loss/reg": 0.0,
+      "step": 10330
+    },
+    {
+      "epoch": 0.06802631578947368,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.0734283447265625,
+      "learning_rate": 0.0001,
+      "loss": 3.2653,
+      "loss/crossentropy": 2.3983714103698732,
+      "loss/hidden": 2.9625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29214349389076233,
+      "loss/reg": 0.0,
+      "step": 10340
+    },
+    {
+      "epoch": 0.0680921052631579,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.0652984619140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2579,
+      "loss/crossentropy": 2.5644800424575807,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3019869774580002,
+      "loss/reg": 0.0,
+      "step": 10350
+    },
+    {
+      "epoch": 0.06815789473684211,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 3.350255903165907e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.3114,
+      "loss/crossentropy": 2.6736939191818236,
+      "loss/hidden": 2.9015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2768147349357605,
+      "loss/reg": 0.0,
+      "step": 10360
+    },
+    {
+      "epoch": 0.06822368421052631,
+      "grad_norm": 4.65625,
+      "grad_norm_var": 3.350255902575034e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.3131,
+      "loss/crossentropy": 2.2038461327552796,
+      "loss/hidden": 3.0859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31388128101825713,
+      "loss/reg": 0.0,
+      "step": 10370
+    },
+    {
+      "epoch": 0.06828947368421052,
+      "grad_norm": 3.78125,
+      "grad_norm_var": 0.5317220052083333,
+      "learning_rate": 0.0001,
+      "loss": 3.3243,
+      "loss/crossentropy": 2.3583734750747682,
+      "loss/hidden": 2.821875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26419220119714737,
+      "loss/reg": 0.0,
+      "step": 10380
+    },
+    {
+      "epoch": 0.06835526315789474,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.2758127848307292,
+      "learning_rate": 0.0001,
+      "loss": 3.287,
+      "loss/crossentropy": 2.2863503098487854,
+      "loss/hidden": 2.846875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23900726288557053,
+      "loss/reg": 0.0,
+      "step": 10390
+    },
+    {
+      "epoch": 0.06842105263157895,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.04855143229166667,
+      "learning_rate": 0.0001,
+      "loss": 3.2604,
+      "loss/crossentropy": 2.3840556263923647,
+      "loss/hidden": 3.14375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30461184978485106,
+      "loss/reg": 0.0,
+      "step": 10400
+    },
+    {
+      "epoch": 0.06848684210526315,
+      "grad_norm": 2.953125,
+      "grad_norm_var": 0.15156962076822916,
+      "learning_rate": 0.0001,
+      "loss": 3.3062,
+      "loss/crossentropy": 2.088348960876465,
+      "loss/hidden": 3.01875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2698833361268044,
+      "loss/reg": 0.0,
+      "step": 10410
+    },
+    {
+      "epoch": 0.06855263157894737,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.5885894775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.3402,
+      "loss/crossentropy": 2.551842737197876,
+      "loss/hidden": 3.0109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.313777893781662,
+      "loss/reg": 0.0,
+      "step": 10420
+    },
+    {
+      "epoch": 0.06861842105263158,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.9581939697265625,
+      "learning_rate": 0.0001,
+      "loss": 3.3486,
+      "loss/crossentropy": 2.148633885383606,
+      "loss/hidden": 3.1921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29070123881101606,
+      "loss/reg": 0.0,
+      "step": 10430
+    },
+    {
+      "epoch": 0.06868421052631579,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.2256988525390625,
+      "learning_rate": 0.0001,
+      "loss": 3.3073,
+      "loss/crossentropy": 2.0771638333797453,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24762868583202363,
+      "loss/reg": 0.0,
+      "step": 10440
+    },
+    {
+      "epoch": 0.06875,
+      "grad_norm": 1.9453125,
+      "grad_norm_var": 0.30576553344726565,
+      "learning_rate": 0.0001,
+      "loss": 3.3608,
+      "loss/crossentropy": 1.9622422456741333,
+      "loss/hidden": 3.046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2823460906744003,
+      "loss/reg": 0.0,
+      "step": 10450
+    },
+    {
+      "epoch": 0.0688157894736842,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.24808731079101562,
+      "learning_rate": 0.0001,
+      "loss": 3.2114,
+      "loss/crossentropy": 2.4866459488868715,
+      "loss/hidden": 2.8640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25186999291181567,
+      "loss/reg": 0.0,
+      "step": 10460
+    },
+    {
+      "epoch": 0.06888157894736842,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 3.232233683268229,
+      "learning_rate": 0.0001,
+      "loss": 3.2672,
+      "loss/crossentropy": 2.4218720883131026,
+      "loss/hidden": 3.2421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34285663813352585,
+      "loss/reg": 0.0,
+      "step": 10470
+    },
+    {
+      "epoch": 0.06894736842105263,
+      "grad_norm": 3.0,
+      "grad_norm_var": 0.12942301432291667,
+      "learning_rate": 0.0001,
+      "loss": 3.2774,
+      "loss/crossentropy": 2.2948715806007387,
+      "loss/hidden": 2.765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22767567485570908,
+      "loss/reg": 0.0,
+      "step": 10480
+    },
+    {
+      "epoch": 0.06901315789473685,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.09120686848958333,
+      "learning_rate": 0.0001,
+      "loss": 3.2795,
+      "loss/crossentropy": 2.4549028277397156,
+      "loss/hidden": 3.0515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2874615803360939,
+      "loss/reg": 0.0,
+      "step": 10490
+    },
+    {
+      "epoch": 0.06907894736842106,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.14446512858072916,
+      "learning_rate": 0.0001,
+      "loss": 3.2562,
+      "loss/crossentropy": 2.2758328318595886,
+      "loss/hidden": 2.846875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24516315162181854,
+      "loss/reg": 0.0,
+      "step": 10500
+    },
+    {
+      "epoch": 0.06914473684210526,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.167626953125,
+      "learning_rate": 0.0001,
+      "loss": 3.3066,
+      "loss/crossentropy": 2.422711133956909,
+      "loss/hidden": 2.8671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24841044396162032,
+      "loss/reg": 0.0,
+      "step": 10510
+    },
+    {
+      "epoch": 0.06921052631578947,
+      "grad_norm": 2.953125,
+      "grad_norm_var": 0.0686187744140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2834,
+      "loss/crossentropy": 2.210281264781952,
+      "loss/hidden": 3.3,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31483527421951296,
+      "loss/reg": 0.0,
+      "step": 10520
+    },
+    {
+      "epoch": 0.06927631578947369,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.1522857666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.2678,
+      "loss/crossentropy": 2.3035526394844057,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26298564970493316,
+      "loss/reg": 0.0,
+      "step": 10530
+    },
+    {
+      "epoch": 0.0693421052631579,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 1.5601959228515625,
+      "learning_rate": 0.0001,
+      "loss": 3.2926,
+      "loss/crossentropy": 2.5075936675071717,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27058843374252317,
+      "loss/reg": 0.0,
+      "step": 10540
+    },
+    {
+      "epoch": 0.0694078947368421,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 1.5597330729166667,
+      "learning_rate": 0.0001,
+      "loss": 3.2643,
+      "loss/crossentropy": 2.448484253883362,
+      "loss/hidden": 3.1421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3164616659283638,
+      "loss/reg": 0.0,
+      "step": 10550
+    },
+    {
+      "epoch": 0.06947368421052631,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.3221638997395833,
+      "learning_rate": 0.0001,
+      "loss": 3.2548,
+      "loss/crossentropy": 2.3997972130775453,
+      "loss/hidden": 3.4125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3671171858906746,
+      "loss/reg": 0.0,
+      "step": 10560
+    },
+    {
+      "epoch": 0.06953947368421053,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.33424072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.3035,
+      "loss/crossentropy": 2.3439933180809023,
+      "loss/hidden": 2.8671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24507830888032914,
+      "loss/reg": 0.0,
+      "step": 10570
+    },
+    {
+      "epoch": 0.06960526315789474,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.15203450520833334,
+      "learning_rate": 0.0001,
+      "loss": 3.3203,
+      "loss/crossentropy": 2.20079083442688,
+      "loss/hidden": 3.203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3732150986790657,
+      "loss/reg": 0.0,
+      "step": 10580
+    },
+    {
+      "epoch": 0.06967105263157895,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.23407796223958333,
+      "learning_rate": 0.0001,
+      "loss": 3.3251,
+      "loss/crossentropy": 2.353682446479797,
+      "loss/hidden": 3.1125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2952089346945286,
+      "loss/reg": 0.0,
+      "step": 10590
+    },
+    {
+      "epoch": 0.06973684210526315,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.22082926432291666,
+      "learning_rate": 0.0001,
+      "loss": 3.2719,
+      "loss/crossentropy": 2.3233517169952393,
+      "loss/hidden": 3.1015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.290130452811718,
+      "loss/reg": 0.0,
+      "step": 10600
+    },
+    {
+      "epoch": 0.06980263157894737,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.2807362874348958,
+      "learning_rate": 0.0001,
+      "loss": 3.2558,
+      "loss/crossentropy": 2.466183233261108,
+      "loss/hidden": 3.096875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33012075573205946,
+      "loss/reg": 0.0,
+      "step": 10610
+    },
+    {
+      "epoch": 0.06986842105263158,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.08990478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.3285,
+      "loss/crossentropy": 2.3989938259124757,
+      "loss/hidden": 3.184375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37008936554193494,
+      "loss/reg": 0.0,
+      "step": 10620
+    },
+    {
+      "epoch": 0.0699342105263158,
+      "grad_norm": 3.09375,
+      "grad_norm_var": 0.05269266764322917,
+      "learning_rate": 0.0001,
+      "loss": 3.2761,
+      "loss/crossentropy": 2.4531158804893494,
+      "loss/hidden": 2.96875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27111856192350386,
+      "loss/reg": 0.0,
+      "step": 10630
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.0650787353515625,
+      "learning_rate": 0.0001,
+      "loss": 3.22,
+      "loss/crossentropy": 2.44055380821228,
+      "loss/hidden": 2.9171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25391832590103147,
+      "loss/reg": 0.0,
+      "step": 10640
+    },
+    {
+      "epoch": 0.0700657894736842,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.0595367431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.226,
+      "loss/crossentropy": 2.4790910482406616,
+      "loss/hidden": 2.975,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2732370227575302,
+      "loss/reg": 0.0,
+      "step": 10650
+    },
+    {
+      "epoch": 0.07013157894736842,
+      "grad_norm": 2248146944.0,
+      "grad_norm_var": 3.158852919285514e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.3733,
+      "loss/crossentropy": 2.1218923926353455,
+      "loss/hidden": 2.7578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23474968373775482,
+      "loss/reg": 0.0,
+      "step": 10660
+    },
+    {
+      "epoch": 0.07019736842105263,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 3.158852918981078e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2364,
+      "loss/crossentropy": 2.3490695118904115,
+      "loss/hidden": 2.984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3323436751961708,
+      "loss/reg": 0.0,
+      "step": 10670
+    },
+    {
+      "epoch": 0.07026315789473685,
+      "grad_norm": 3.1875,
+      "grad_norm_var": 0.07629801432291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1831,
+      "loss/crossentropy": 2.3577569365501403,
+      "loss/hidden": 2.90625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26663027703762054,
+      "loss/reg": 0.0,
+      "step": 10680
+    },
+    {
+      "epoch": 0.07032894736842105,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.0871978759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.2111,
+      "loss/crossentropy": 2.2620568752288817,
+      "loss/hidden": 3.0796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3401878133416176,
+      "loss/reg": 0.0,
+      "step": 10690
+    },
+    {
+      "epoch": 0.07039473684210526,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.03775634765625,
+      "learning_rate": 0.0001,
+      "loss": 3.252,
+      "loss/crossentropy": 2.223601281642914,
+      "loss/hidden": 2.9671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26251988410949706,
+      "loss/reg": 0.0,
+      "step": 10700
+    },
+    {
+      "epoch": 0.07046052631578947,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.020702107747395834,
+      "learning_rate": 0.0001,
+      "loss": 3.1904,
+      "loss/crossentropy": 2.2720033645629885,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2705120757222176,
+      "loss/reg": 0.0,
+      "step": 10710
+    },
+    {
+      "epoch": 0.07052631578947369,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.36824442545572916,
+      "learning_rate": 0.0001,
+      "loss": 3.2981,
+      "loss/crossentropy": 2.3247862100601195,
+      "loss/hidden": 2.8796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25748861730098727,
+      "loss/reg": 0.0,
+      "step": 10720
+    },
+    {
+      "epoch": 0.0705921052631579,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.27925796508789064,
+      "learning_rate": 0.0001,
+      "loss": 3.1731,
+      "loss/crossentropy": 1.9388428241014481,
+      "loss/hidden": 2.828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21091363951563835,
+      "loss/reg": 0.0,
+      "step": 10730
+    },
+    {
+      "epoch": 0.0706578947368421,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.10903294881184895,
+      "learning_rate": 0.0001,
+      "loss": 3.2623,
+      "loss/crossentropy": 2.338471806049347,
+      "loss/hidden": 2.8875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2822930008172989,
+      "loss/reg": 0.0,
+      "step": 10740
+    },
+    {
+      "epoch": 0.07072368421052631,
+      "grad_norm": 1.96875,
+      "grad_norm_var": 0.11550191243489584,
+      "learning_rate": 0.0001,
+      "loss": 3.2306,
+      "loss/crossentropy": 2.275705647468567,
+      "loss/hidden": 3.1640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30476620346307753,
+      "loss/reg": 0.0,
+      "step": 10750
+    },
+    {
+      "epoch": 0.07078947368421053,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 1.7166033426920573,
+      "learning_rate": 0.0001,
+      "loss": 3.1802,
+      "loss/crossentropy": 2.375147843360901,
+      "loss/hidden": 2.8625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3041233107447624,
+      "loss/reg": 0.0,
+      "step": 10760
+    },
+    {
+      "epoch": 0.07085526315789474,
+      "grad_norm": 2.75,
+      "grad_norm_var": 1.698127237955729,
+      "learning_rate": 0.0001,
+      "loss": 3.308,
+      "loss/crossentropy": 2.3641104817390444,
+      "loss/hidden": 2.9734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2578106954693794,
+      "loss/reg": 0.0,
+      "step": 10770
+    },
+    {
+      "epoch": 0.07092105263157895,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.33463134765625,
+      "learning_rate": 0.0001,
+      "loss": 3.2267,
+      "loss/crossentropy": 2.3689509868621825,
+      "loss/hidden": 3.053125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3170511037111282,
+      "loss/reg": 0.0,
+      "step": 10780
+    },
+    {
+      "epoch": 0.07098684210526315,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.08876851399739584,
+      "learning_rate": 0.0001,
+      "loss": 3.2416,
+      "loss/crossentropy": 2.0522693753242494,
+      "loss/hidden": 2.884375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25120987445116044,
+      "loss/reg": 0.0,
+      "step": 10790
+    },
+    {
+      "epoch": 0.07105263157894737,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.1547027587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.2808,
+      "loss/crossentropy": 2.193123185634613,
+      "loss/hidden": 2.94375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30211101323366163,
+      "loss/reg": 0.0,
+      "step": 10800
+    },
+    {
+      "epoch": 0.07111842105263158,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.11612955729166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1868,
+      "loss/crossentropy": 2.6722516775131226,
+      "loss/hidden": 2.86875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26585151851177213,
+      "loss/reg": 0.0,
+      "step": 10810
+    },
+    {
+      "epoch": 0.0711842105263158,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.3016998291015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1683,
+      "loss/crossentropy": 2.480617439746857,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2681833073496819,
+      "loss/reg": 0.0,
+      "step": 10820
+    },
+    {
+      "epoch": 0.07125,
+      "grad_norm": 3.8125,
+      "grad_norm_var": 0.1811431884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.2239,
+      "loss/crossentropy": 2.2105371236801146,
+      "loss/hidden": 2.8875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27378717362880706,
+      "loss/reg": 0.0,
+      "step": 10830
+    },
+    {
+      "epoch": 0.07131578947368421,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.17531636555989583,
+      "learning_rate": 0.0001,
+      "loss": 3.1981,
+      "loss/crossentropy": 2.201746928691864,
+      "loss/hidden": 3.2078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3672773316502571,
+      "loss/reg": 0.0,
+      "step": 10840
+    },
+    {
+      "epoch": 0.07138157894736842,
+      "grad_norm": 4.28125,
+      "grad_norm_var": 0.34869791666666666,
+      "learning_rate": 0.0001,
+      "loss": 3.2799,
+      "loss/crossentropy": 2.322747588157654,
+      "loss/hidden": 2.9625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.262615929543972,
+      "loss/reg": 0.0,
+      "step": 10850
+    },
+    {
+      "epoch": 0.07144736842105263,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.7465810139973958,
+      "learning_rate": 0.0001,
+      "loss": 3.2065,
+      "loss/crossentropy": 1.9148864209651948,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2080842524766922,
+      "loss/reg": 0.0,
+      "step": 10860
+    },
+    {
+      "epoch": 0.07151315789473685,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.053873697916666664,
+      "learning_rate": 0.0001,
+      "loss": 3.2002,
+      "loss/crossentropy": 2.314997375011444,
+      "loss/hidden": 2.9234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2829948261380196,
+      "loss/reg": 0.0,
+      "step": 10870
+    },
+    {
+      "epoch": 0.07157894736842105,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.08782145182291666,
+      "learning_rate": 0.0001,
+      "loss": 3.2222,
+      "loss/crossentropy": 2.4381492733955383,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2620039567351341,
+      "loss/reg": 0.0,
+      "step": 10880
+    },
+    {
+      "epoch": 0.07164473684210526,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 0.10087788899739583,
+      "learning_rate": 0.0001,
+      "loss": 3.2769,
+      "loss/crossentropy": 2.478635573387146,
+      "loss/hidden": 3.028125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30705118626356126,
+      "loss/reg": 0.0,
+      "step": 10890
+    },
+    {
+      "epoch": 0.07171052631578947,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.12694066365559895,
+      "learning_rate": 0.0001,
+      "loss": 3.1767,
+      "loss/crossentropy": 2.145402270555496,
+      "loss/hidden": 3.028125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26391510516405103,
+      "loss/reg": 0.0,
+      "step": 10900
+    },
+    {
+      "epoch": 0.07177631578947369,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.13940404256184896,
+      "learning_rate": 0.0001,
+      "loss": 3.2976,
+      "loss/crossentropy": 2.1119534373283386,
+      "loss/hidden": 3.028125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24341508150100707,
+      "loss/reg": 0.0,
+      "step": 10910
+    },
+    {
+      "epoch": 0.0718421052631579,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.193408203125,
+      "learning_rate": 0.0001,
+      "loss": 3.2699,
+      "loss/crossentropy": 2.2503302097320557,
+      "loss/hidden": 3.196875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2697600871324539,
+      "loss/reg": 0.0,
+      "step": 10920
+    },
+    {
+      "epoch": 0.0719078947368421,
+      "grad_norm": 4.9375,
+      "grad_norm_var": 1.1786092122395833,
+      "learning_rate": 0.0001,
+      "loss": 3.3635,
+      "loss/crossentropy": 2.2035016298294066,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23886747062206268,
+      "loss/reg": 0.0,
+      "step": 10930
+    },
+    {
+      "epoch": 0.07197368421052631,
+      "grad_norm": 2.5,
+      "grad_norm_var": 1.1909088134765624,
+      "learning_rate": 0.0001,
+      "loss": 3.2664,
+      "loss/crossentropy": 2.438611125946045,
+      "loss/hidden": 2.9390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2649322673678398,
+      "loss/reg": 0.0,
+      "step": 10940
+    },
+    {
+      "epoch": 0.07203947368421053,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.1803375244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2795,
+      "loss/crossentropy": 2.300194537639618,
+      "loss/hidden": 2.9546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2784364491701126,
+      "loss/reg": 0.0,
+      "step": 10950
+    },
+    {
+      "epoch": 0.07210526315789474,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.13961181640625,
+      "learning_rate": 0.0001,
+      "loss": 3.196,
+      "loss/crossentropy": 2.297783041000366,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23439482748508453,
+      "loss/reg": 0.0,
+      "step": 10960
+    },
+    {
+      "epoch": 0.07217105263157894,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.03483784993489583,
+      "learning_rate": 0.0001,
+      "loss": 3.2816,
+      "loss/crossentropy": 2.2468614101409914,
+      "loss/hidden": 3.0,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26336451917886733,
+      "loss/reg": 0.0,
+      "step": 10970
+    },
+    {
+      "epoch": 0.07223684210526315,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.027684529622395832,
+      "learning_rate": 0.0001,
+      "loss": 3.2387,
+      "loss/crossentropy": 2.5107889652252195,
+      "loss/hidden": 2.8984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2557321682572365,
+      "loss/reg": 0.0,
+      "step": 10980
+    },
+    {
+      "epoch": 0.07230263157894737,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.05127665201822917,
+      "learning_rate": 0.0001,
+      "loss": 3.2995,
+      "loss/crossentropy": 2.221148931980133,
+      "loss/hidden": 2.9859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2837982401251793,
+      "loss/reg": 0.0,
+      "step": 10990
+    },
+    {
+      "epoch": 0.07236842105263158,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.08108317057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.289,
+      "loss/crossentropy": 2.515943694114685,
+      "loss/hidden": 2.8890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.269567608833313,
+      "loss/reg": 0.0,
+      "step": 11000
+    },
+    {
+      "epoch": 0.0724342105263158,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.10301106770833333,
+      "learning_rate": 0.0001,
+      "loss": 3.2421,
+      "loss/crossentropy": 2.22921404838562,
+      "loss/hidden": 2.8359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24364713877439498,
+      "loss/reg": 0.0,
+      "step": 11010
+    },
+    {
+      "epoch": 0.0725,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.15533447265625,
+      "learning_rate": 0.0001,
+      "loss": 3.2872,
+      "loss/crossentropy": 2.5498072266578675,
+      "loss/hidden": 2.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2900205120444298,
+      "loss/reg": 0.0,
+      "step": 11020
+    },
+    {
+      "epoch": 0.07256578947368421,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.19057515462239583,
+      "learning_rate": 0.0001,
+      "loss": 3.2853,
+      "loss/crossentropy": 2.607512426376343,
+      "loss/hidden": 3.04375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3007731422781944,
+      "loss/reg": 0.0,
+      "step": 11030
+    },
+    {
+      "epoch": 0.07263157894736842,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.13447240193684895,
+      "learning_rate": 0.0001,
+      "loss": 3.1687,
+      "loss/crossentropy": 1.9578089714050293,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23623087108135224,
+      "loss/reg": 0.0,
+      "step": 11040
+    },
+    {
+      "epoch": 0.07269736842105264,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.2533192952473958,
+      "learning_rate": 0.0001,
+      "loss": 3.2864,
+      "loss/crossentropy": 2.473111295700073,
+      "loss/hidden": 3.334375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31404276490211486,
+      "loss/reg": 0.0,
+      "step": 11050
+    },
+    {
+      "epoch": 0.07276315789473685,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.1464019775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.2587,
+      "loss/crossentropy": 2.1964801430702208,
+      "loss/hidden": 2.9078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26576130986213686,
+      "loss/reg": 0.0,
+      "step": 11060
+    },
+    {
+      "epoch": 0.07282894736842105,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.12649637858072918,
+      "learning_rate": 0.0001,
+      "loss": 3.2663,
+      "loss/crossentropy": 2.2337945103645325,
+      "loss/hidden": 2.8984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2691087871789932,
+      "loss/reg": 0.0,
+      "step": 11070
+    },
+    {
+      "epoch": 0.07289473684210526,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.08884175618489583,
+      "learning_rate": 0.0001,
+      "loss": 3.3071,
+      "loss/crossentropy": 2.5003953099250795,
+      "loss/hidden": 3.021875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3058940455317497,
+      "loss/reg": 0.0,
+      "step": 11080
+    },
+    {
+      "epoch": 0.07296052631578948,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.14674479166666668,
+      "learning_rate": 0.0001,
+      "loss": 3.2275,
+      "loss/crossentropy": 2.3431849002838137,
+      "loss/hidden": 3.003125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2692634254693985,
+      "loss/reg": 0.0,
+      "step": 11090
+    },
+    {
+      "epoch": 0.07302631578947369,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.08789774576822916,
+      "learning_rate": 0.0001,
+      "loss": 3.2168,
+      "loss/crossentropy": 2.2301442503929136,
+      "loss/hidden": 2.8625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24614981114864348,
+      "loss/reg": 0.0,
+      "step": 11100
+    },
+    {
+      "epoch": 0.07309210526315789,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.09053446451822916,
+      "learning_rate": 0.0001,
+      "loss": 3.3336,
+      "loss/crossentropy": 2.624728870391846,
+      "loss/hidden": 2.925,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30191341042518616,
+      "loss/reg": 0.0,
+      "step": 11110
+    },
+    {
+      "epoch": 0.0731578947368421,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.23191731770833332,
+      "learning_rate": 0.0001,
+      "loss": 3.3217,
+      "loss/crossentropy": 2.204717183113098,
+      "loss/hidden": 3.4859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.5070606812834739,
+      "loss/reg": 0.0,
+      "step": 11120
+    },
+    {
+      "epoch": 0.07322368421052632,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.2526519775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.345,
+      "loss/crossentropy": 2.558793139457703,
+      "loss/hidden": 3.0828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.42952366173267365,
+      "loss/reg": 0.0,
+      "step": 11130
+    },
+    {
+      "epoch": 0.07328947368421053,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.0856109619140625,
+      "learning_rate": 0.0001,
+      "loss": 3.3754,
+      "loss/crossentropy": 2.227391791343689,
+      "loss/hidden": 2.8890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27742871195077895,
+      "loss/reg": 0.0,
+      "step": 11140
+    },
+    {
+      "epoch": 0.07335526315789474,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.07666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.2194,
+      "loss/crossentropy": 2.32956976890564,
+      "loss/hidden": 2.9609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2826590985059738,
+      "loss/reg": 0.0,
+      "step": 11150
+    },
+    {
+      "epoch": 0.07342105263157894,
+      "grad_norm": 8.625,
+      "grad_norm_var": 2.3614542643229166,
+      "learning_rate": 0.0001,
+      "loss": 3.2431,
+      "loss/crossentropy": 2.0696181058883667,
+      "loss/hidden": 3.3640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2718909472227097,
+      "loss/reg": 0.0,
+      "step": 11160
+    },
+    {
+      "epoch": 0.07348684210526316,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 2.3623697916666666,
+      "learning_rate": 0.0001,
+      "loss": 3.2983,
+      "loss/crossentropy": 2.5612054228782655,
+      "loss/hidden": 2.94375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26728835999965667,
+      "loss/reg": 0.0,
+      "step": 11170
+    },
+    {
+      "epoch": 0.07355263157894737,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.11398111979166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1779,
+      "loss/crossentropy": 2.5427613735198973,
+      "loss/hidden": 2.775,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24258261919021606,
+      "loss/reg": 0.0,
+      "step": 11180
+    },
+    {
+      "epoch": 0.07361842105263158,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.07968648274739583,
+      "learning_rate": 0.0001,
+      "loss": 3.1565,
+      "loss/crossentropy": 2.3772116184234617,
+      "loss/hidden": 2.734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24017660170793534,
+      "loss/reg": 0.0,
+      "step": 11190
+    },
+    {
+      "epoch": 0.07368421052631578,
+      "grad_norm": 2.9375,
+      "grad_norm_var": 0.058652496337890624,
+      "learning_rate": 0.0001,
+      "loss": 3.2383,
+      "loss/crossentropy": 2.284471166133881,
+      "loss/hidden": 3.0625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31355464905500413,
+      "loss/reg": 0.0,
+      "step": 11200
+    },
+    {
+      "epoch": 0.07375,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.04164937337239583,
+      "learning_rate": 0.0001,
+      "loss": 3.2288,
+      "loss/crossentropy": 2.1727640271186828,
+      "loss/hidden": 2.815625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24059856683015823,
+      "loss/reg": 0.0,
+      "step": 11210
+    },
+    {
+      "epoch": 0.07381578947368421,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.015363566080729167,
+      "learning_rate": 0.0001,
+      "loss": 3.2857,
+      "loss/crossentropy": 2.377478325366974,
+      "loss/hidden": 3.0984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2916824325919151,
+      "loss/reg": 0.0,
+      "step": 11220
+    },
+    {
+      "epoch": 0.07388157894736842,
+      "grad_norm": 3.140625,
+      "grad_norm_var": 0.04091695149739583,
+      "learning_rate": 0.0001,
+      "loss": 3.2848,
+      "loss/crossentropy": 2.372981941699982,
+      "loss/hidden": 3.140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3395596519112587,
+      "loss/reg": 0.0,
+      "step": 11230
+    },
+    {
+      "epoch": 0.07394736842105264,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 1.1325103759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.2954,
+      "loss/crossentropy": 2.571339511871338,
+      "loss/hidden": 2.9953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3173587560653687,
+      "loss/reg": 0.0,
+      "step": 11240
+    },
+    {
+      "epoch": 0.07401315789473684,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.11298421223958334,
+      "learning_rate": 0.0001,
+      "loss": 3.213,
+      "loss/crossentropy": 2.093873751163483,
+      "loss/hidden": 2.8875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23975073918700218,
+      "loss/reg": 0.0,
+      "step": 11250
+    },
+    {
+      "epoch": 0.07407894736842105,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.13909403483072916,
+      "learning_rate": 0.0001,
+      "loss": 3.25,
+      "loss/crossentropy": 2.55220787525177,
+      "loss/hidden": 3.00625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29909972846508026,
+      "loss/reg": 0.0,
+      "step": 11260
+    },
+    {
+      "epoch": 0.07414473684210526,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.11393229166666667,
+      "learning_rate": 0.0001,
+      "loss": 3.2424,
+      "loss/crossentropy": 2.18590772151947,
+      "loss/hidden": 2.98125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2539385199546814,
+      "loss/reg": 0.0,
+      "step": 11270
+    },
+    {
+      "epoch": 0.07421052631578948,
+      "grad_norm": 3.078125,
+      "grad_norm_var": 0.07273661295572917,
+      "learning_rate": 0.0001,
+      "loss": 3.2354,
+      "loss/crossentropy": 2.4804351210594175,
+      "loss/hidden": 2.8953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2747698500752449,
+      "loss/reg": 0.0,
+      "step": 11280
+    },
+    {
+      "epoch": 0.07427631578947369,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.17068684895833333,
+      "learning_rate": 0.0001,
+      "loss": 3.2114,
+      "loss/crossentropy": 2.4408376574516297,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3068179443478584,
+      "loss/reg": 0.0,
+      "step": 11290
+    },
+    {
+      "epoch": 0.07434210526315789,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.27847900390625,
+      "learning_rate": 0.0001,
+      "loss": 3.3225,
+      "loss/crossentropy": 2.284189748764038,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2627576723694801,
+      "loss/reg": 0.0,
+      "step": 11300
+    },
+    {
+      "epoch": 0.0744078947368421,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.22431233723958333,
+      "learning_rate": 0.0001,
+      "loss": 3.2794,
+      "loss/crossentropy": 2.348969095945358,
+      "loss/hidden": 2.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22996244430541993,
+      "loss/reg": 0.0,
+      "step": 11310
+    },
+    {
+      "epoch": 0.07447368421052632,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.8210245768229166,
+      "learning_rate": 0.0001,
+      "loss": 3.2548,
+      "loss/crossentropy": 2.455811655521393,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2752078205347061,
+      "loss/reg": 0.0,
+      "step": 11320
+    },
+    {
+      "epoch": 0.07453947368421053,
+      "grad_norm": 4.75,
+      "grad_norm_var": 0.3614095052083333,
+      "learning_rate": 0.0001,
+      "loss": 3.2713,
+      "loss/crossentropy": 2.4881197214126587,
+      "loss/hidden": 3.1234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33918842375278474,
+      "loss/reg": 0.0,
+      "step": 11330
+    },
+    {
+      "epoch": 0.07460526315789473,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.4044596354166667,
+      "learning_rate": 0.0001,
+      "loss": 3.334,
+      "loss/crossentropy": 2.4584757328033446,
+      "loss/hidden": 3.0046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3295674562454224,
+      "loss/reg": 0.0,
+      "step": 11340
+    },
+    {
+      "epoch": 0.07467105263157894,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.9745920817057292,
+      "learning_rate": 0.0001,
+      "loss": 3.2347,
+      "loss/crossentropy": 2.413297247886658,
+      "loss/hidden": 3.0890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2875028237700462,
+      "loss/reg": 0.0,
+      "step": 11350
+    },
+    {
+      "epoch": 0.07473684210526316,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.0769439697265625,
+      "learning_rate": 0.0001,
+      "loss": 3.2495,
+      "loss/crossentropy": 2.297819769382477,
+      "loss/hidden": 3.0546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28927008211612704,
+      "loss/reg": 0.0,
+      "step": 11360
+    },
+    {
+      "epoch": 0.07480263157894737,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.054182942708333334,
+      "learning_rate": 0.0001,
+      "loss": 3.1982,
+      "loss/crossentropy": 2.2709707379341126,
+      "loss/hidden": 3.03125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29085248410701753,
+      "loss/reg": 0.0,
+      "step": 11370
+    },
+    {
+      "epoch": 0.07486842105263158,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.315283203125,
+      "learning_rate": 0.0001,
+      "loss": 3.2877,
+      "loss/crossentropy": 2.5602762937545775,
+      "loss/hidden": 3.1828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3155085578560829,
+      "loss/reg": 0.0,
+      "step": 11380
+    },
+    {
+      "epoch": 0.07493421052631578,
+      "grad_norm": 8.375,
+      "grad_norm_var": 2.533186848958333,
+      "learning_rate": 0.0001,
+      "loss": 3.2192,
+      "loss/crossentropy": 2.231611895561218,
+      "loss/hidden": 3.009375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27611204236745834,
+      "loss/reg": 0.0,
+      "step": 11390
+    },
+    {
+      "epoch": 0.075,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 2.313280232747396,
+      "learning_rate": 0.0001,
+      "loss": 3.3273,
+      "loss/crossentropy": 2.2164941787719727,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2650335058569908,
+      "loss/reg": 0.0,
+      "step": 11400
+    },
+    {
+      "epoch": 0.07506578947368421,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.3658528645833333,
+      "learning_rate": 0.0001,
+      "loss": 3.2587,
+      "loss/crossentropy": 2.2440002799034118,
+      "loss/hidden": 3.075,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3373932957649231,
+      "loss/reg": 0.0,
+      "step": 11410
+    },
+    {
+      "epoch": 0.07513157894736842,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.06290690104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1439,
+      "loss/crossentropy": 2.405060076713562,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26099992990493776,
+      "loss/reg": 0.0,
+      "step": 11420
+    },
+    {
+      "epoch": 0.07519736842105264,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.04853108723958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1454,
+      "loss/crossentropy": 2.5482593178749084,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2424457401037216,
+      "loss/reg": 0.0,
+      "step": 11430
+    },
+    {
+      "epoch": 0.07526315789473684,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.5184244791666667,
+      "learning_rate": 0.0001,
+      "loss": 3.2189,
+      "loss/crossentropy": 2.1846509099006655,
+      "loss/hidden": 3.0390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29890656769275664,
+      "loss/reg": 0.0,
+      "step": 11440
+    },
+    {
+      "epoch": 0.07532894736842105,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.03676656087239583,
+      "learning_rate": 0.0001,
+      "loss": 3.2528,
+      "loss/crossentropy": 2.3350785970687866,
+      "loss/hidden": 3.0015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26758097261190417,
+      "loss/reg": 0.0,
+      "step": 11450
+    },
+    {
+      "epoch": 0.07539473684210526,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 56.07302958170573,
+      "learning_rate": 0.0001,
+      "loss": 3.3264,
+      "loss/crossentropy": 2.357296335697174,
+      "loss/hidden": 3.0171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32786626666784285,
+      "loss/reg": 0.0,
+      "step": 11460
+    },
+    {
+      "epoch": 0.07546052631578948,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.05142822265625,
+      "learning_rate": 0.0001,
+      "loss": 3.1956,
+      "loss/crossentropy": 2.29429577589035,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22037848085165024,
+      "loss/reg": 0.0,
+      "step": 11470
+    },
+    {
+      "epoch": 0.07552631578947368,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.09621988932291667,
+      "learning_rate": 0.0001,
+      "loss": 3.2053,
+      "loss/crossentropy": 2.450187027454376,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27029853165149687,
+      "loss/reg": 0.0,
+      "step": 11480
+    },
+    {
+      "epoch": 0.07559210526315789,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.06201171875,
+      "learning_rate": 0.0001,
+      "loss": 3.2252,
+      "loss/crossentropy": 1.9360981225967406,
+      "loss/hidden": 2.9296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23251112401485444,
+      "loss/reg": 0.0,
+      "step": 11490
+    },
+    {
+      "epoch": 0.0756578947368421,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.07566731770833333,
+      "learning_rate": 0.0001,
+      "loss": 3.2003,
+      "loss/crossentropy": 2.21165417432785,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2515560120344162,
+      "loss/reg": 0.0,
+      "step": 11500
+    },
+    {
+      "epoch": 0.07572368421052632,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.07636311848958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1758,
+      "loss/crossentropy": 2.5108731746673585,
+      "loss/hidden": 3.0515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31811543107032775,
+      "loss/reg": 0.0,
+      "step": 11510
+    },
+    {
+      "epoch": 0.07578947368421053,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.17082926432291667,
+      "learning_rate": 0.0001,
+      "loss": 3.2811,
+      "loss/crossentropy": 2.1942604899406435,
+      "loss/hidden": 2.8453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25081480443477633,
+      "loss/reg": 0.0,
+      "step": 11520
+    },
+    {
+      "epoch": 0.07585526315789473,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.07385660807291666,
+      "learning_rate": 0.0001,
+      "loss": 3.239,
+      "loss/crossentropy": 2.0190611362457274,
+      "loss/hidden": 3.00625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28278502970933916,
+      "loss/reg": 0.0,
+      "step": 11530
+    },
+    {
+      "epoch": 0.07592105263157894,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.045166015625,
+      "learning_rate": 0.0001,
+      "loss": 3.2825,
+      "loss/crossentropy": 2.3928887605667115,
+      "loss/hidden": 3.0375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31386475563049315,
+      "loss/reg": 0.0,
+      "step": 11540
+    },
+    {
+      "epoch": 0.07598684210526316,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.05419514973958333,
+      "learning_rate": 0.0001,
+      "loss": 3.236,
+      "loss/crossentropy": 2.3381729245185854,
+      "loss/hidden": 2.846875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2617738708853722,
+      "loss/reg": 0.0,
+      "step": 11550
+    },
+    {
+      "epoch": 0.07605263157894737,
+      "grad_norm": 3.421875,
+      "grad_norm_var": 0.10877278645833334,
+      "learning_rate": 0.0001,
+      "loss": 3.1435,
+      "loss/crossentropy": 2.1223382353782654,
+      "loss/hidden": 2.9671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2554148808121681,
+      "loss/reg": 0.0,
+      "step": 11560
+    },
+    {
+      "epoch": 0.07611842105263159,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.13093973795572916,
+      "learning_rate": 0.0001,
+      "loss": 3.2511,
+      "loss/crossentropy": 2.4102694511413576,
+      "loss/hidden": 3.0125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2813078135251999,
+      "loss/reg": 0.0,
+      "step": 11570
+    },
+    {
+      "epoch": 0.07618421052631578,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.12099507649739584,
+      "learning_rate": 0.0001,
+      "loss": 3.2712,
+      "loss/crossentropy": 2.2883424520492555,
+      "loss/hidden": 2.9828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27575887441635133,
+      "loss/reg": 0.0,
+      "step": 11580
+    },
+    {
+      "epoch": 0.07625,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.15095926920572916,
+      "learning_rate": 0.0001,
+      "loss": 3.2856,
+      "loss/crossentropy": 2.4679728865623476,
+      "loss/hidden": 2.990625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3342022061347961,
+      "loss/reg": 0.0,
+      "step": 11590
+    },
+    {
+      "epoch": 0.07631578947368421,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.1062896728515625,
+      "learning_rate": 0.0001,
+      "loss": 3.2302,
+      "loss/crossentropy": 2.35604043006897,
+      "loss/hidden": 2.95,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30481296926736834,
+      "loss/reg": 0.0,
+      "step": 11600
+    },
+    {
+      "epoch": 0.07638157894736843,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.17685445149739584,
+      "learning_rate": 0.0001,
+      "loss": 3.3621,
+      "loss/crossentropy": 2.302362835407257,
+      "loss/hidden": 3.153125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29829359203577044,
+      "loss/reg": 0.0,
+      "step": 11610
+    },
+    {
+      "epoch": 0.07644736842105262,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 2.837443679954338e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.365,
+      "loss/crossentropy": 2.3786125659942625,
+      "loss/hidden": 2.9078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2504860758781433,
+      "loss/reg": 0.0,
+      "step": 11620
+    },
+    {
+      "epoch": 0.07651315789473684,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 2.8374436804315274e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.285,
+      "loss/crossentropy": 1.7914829134941102,
+      "loss/hidden": 2.9625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26675148904323576,
+      "loss/reg": 0.0,
+      "step": 11630
+    },
+    {
+      "epoch": 0.07657894736842105,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.047028605143229166,
+      "learning_rate": 0.0001,
+      "loss": 3.2398,
+      "loss/crossentropy": 2.3633928418159487,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2506480649113655,
+      "loss/reg": 0.0,
+      "step": 11640
+    },
+    {
+      "epoch": 0.07664473684210527,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.12795308430989583,
+      "learning_rate": 0.0001,
+      "loss": 3.1729,
+      "loss/crossentropy": 2.3739442467689513,
+      "loss/hidden": 2.8640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24874649345874786,
+      "loss/reg": 0.0,
+      "step": 11650
+    },
+    {
+      "epoch": 0.07671052631578948,
+      "grad_norm": 3.140625,
+      "grad_norm_var": 0.057938639322916666,
+      "learning_rate": 0.0001,
+      "loss": 3.19,
+      "loss/crossentropy": 1.946711039543152,
+      "loss/hidden": 2.946875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26388829201459885,
+      "loss/reg": 0.0,
+      "step": 11660
+    },
+    {
+      "epoch": 0.07677631578947368,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.16907145182291666,
+      "learning_rate": 0.0001,
+      "loss": 3.2141,
+      "loss/crossentropy": 2.5971063375473022,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29624524116516116,
+      "loss/reg": 0.0,
+      "step": 11670
+    },
+    {
+      "epoch": 0.07684210526315789,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.20071512858072918,
+      "learning_rate": 0.0001,
+      "loss": 3.2566,
+      "loss/crossentropy": 2.5601096868515016,
+      "loss/hidden": 2.8890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.272810535132885,
+      "loss/reg": 0.0,
+      "step": 11680
+    },
+    {
+      "epoch": 0.0769078947368421,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.14016927083333333,
+      "learning_rate": 0.0001,
+      "loss": 3.1653,
+      "loss/crossentropy": 2.4755476355552672,
+      "loss/hidden": 2.884375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27636886537075045,
+      "loss/reg": 0.0,
+      "step": 11690
+    },
+    {
+      "epoch": 0.07697368421052632,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.641162109375,
+      "learning_rate": 0.0001,
+      "loss": 3.1798,
+      "loss/crossentropy": 2.558279812335968,
+      "loss/hidden": 2.99375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24465988874435424,
+      "loss/reg": 0.0,
+      "step": 11700
+    },
+    {
+      "epoch": 0.07703947368421053,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.0510162353515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1825,
+      "loss/crossentropy": 2.437064230442047,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25304732471704483,
+      "loss/reg": 0.0,
+      "step": 11710
+    },
+    {
+      "epoch": 0.07710526315789473,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.012482706705729167,
+      "learning_rate": 0.0001,
+      "loss": 3.1497,
+      "loss/crossentropy": 2.3207133412361145,
+      "loss/hidden": 3.0640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26052851378917696,
+      "loss/reg": 0.0,
+      "step": 11720
+    },
+    {
+      "epoch": 0.07717105263157895,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.03972066243489583,
+      "learning_rate": 0.0001,
+      "loss": 3.2354,
+      "loss/crossentropy": 2.210064744949341,
+      "loss/hidden": 3.1765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.321417099237442,
+      "loss/reg": 0.0,
+      "step": 11730
+    },
+    {
+      "epoch": 0.07723684210526316,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.11972249348958333,
+      "learning_rate": 0.0001,
+      "loss": 3.2828,
+      "loss/crossentropy": 2.2914742827415466,
+      "loss/hidden": 3.003125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30591298937797545,
+      "loss/reg": 0.0,
+      "step": 11740
+    },
+    {
+      "epoch": 0.07730263157894737,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.48640034993489584,
+      "learning_rate": 0.0001,
+      "loss": 3.2259,
+      "loss/crossentropy": 2.2794241905212402,
+      "loss/hidden": 3.0546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3180560126900673,
+      "loss/reg": 0.0,
+      "step": 11750
+    },
+    {
+      "epoch": 0.07736842105263157,
+      "grad_norm": 3.15625,
+      "grad_norm_var": 0.4784576416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1508,
+      "loss/crossentropy": 2.237711024284363,
+      "loss/hidden": 2.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22340844720602035,
+      "loss/reg": 0.0,
+      "step": 11760
+    },
+    {
+      "epoch": 0.07743421052631579,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.2001129150390625,
+      "learning_rate": 0.0001,
+      "loss": 3.2992,
+      "loss/crossentropy": 2.4300220131874086,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.279655484855175,
+      "loss/reg": 0.0,
+      "step": 11770
+    },
+    {
+      "epoch": 0.0775,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.0681549072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.1433,
+      "loss/crossentropy": 2.1450002193450928,
+      "loss/hidden": 2.9484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26157657504081727,
+      "loss/reg": 0.0,
+      "step": 11780
+    },
+    {
+      "epoch": 0.07756578947368421,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.038492838541666664,
+      "learning_rate": 0.0001,
+      "loss": 3.2329,
+      "loss/crossentropy": 2.273455095291138,
+      "loss/hidden": 2.921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2625778928399086,
+      "loss/reg": 0.0,
+      "step": 11790
+    },
+    {
+      "epoch": 0.07763157894736843,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.06272379557291667,
+      "learning_rate": 0.0001,
+      "loss": 3.2736,
+      "loss/crossentropy": 2.2713128685951234,
+      "loss/hidden": 2.9421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2852136388421059,
+      "loss/reg": 0.0,
+      "step": 11800
+    },
+    {
+      "epoch": 0.07769736842105263,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.14368082682291666,
+      "learning_rate": 0.0001,
+      "loss": 3.2063,
+      "loss/crossentropy": 2.276504385471344,
+      "loss/hidden": 2.84375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2478427141904831,
+      "loss/reg": 0.0,
+      "step": 11810
+    },
+    {
+      "epoch": 0.07776315789473684,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.14405008951822917,
+      "learning_rate": 0.0001,
+      "loss": 3.2627,
+      "loss/crossentropy": 2.3879762291908264,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26758207380771637,
+      "loss/reg": 0.0,
+      "step": 11820
+    },
+    {
+      "epoch": 0.07782894736842105,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 16.03980712890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1689,
+      "loss/crossentropy": 2.238702917098999,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25569620728492737,
+      "loss/reg": 0.0,
+      "step": 11830
+    },
+    {
+      "epoch": 0.07789473684210527,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 16.03136774698893,
+      "learning_rate": 0.0001,
+      "loss": 3.1737,
+      "loss/crossentropy": 2.11713285446167,
+      "loss/hidden": 2.984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2322181984782219,
+      "loss/reg": 0.0,
+      "step": 11840
+    },
+    {
+      "epoch": 0.07796052631578948,
+      "grad_norm": 3.53125,
+      "grad_norm_var": 0.14957249959309896,
+      "learning_rate": 0.0001,
+      "loss": 3.2023,
+      "loss/crossentropy": 2.279906690120697,
+      "loss/hidden": 2.984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26055515706539156,
+      "loss/reg": 0.0,
+      "step": 11850
+    },
+    {
+      "epoch": 0.07802631578947368,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.1443756103515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1657,
+      "loss/crossentropy": 2.338582932949066,
+      "loss/hidden": 2.815625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23672561049461366,
+      "loss/reg": 0.0,
+      "step": 11860
+    },
+    {
+      "epoch": 0.0780921052631579,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.08467508951822916,
+      "learning_rate": 0.0001,
+      "loss": 3.1991,
+      "loss/crossentropy": 2.3666534066200255,
+      "loss/hidden": 2.9046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27455085664987566,
+      "loss/reg": 0.0,
+      "step": 11870
+    },
+    {
+      "epoch": 0.0781578947368421,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.08329671223958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1862,
+      "loss/crossentropy": 2.232216811180115,
+      "loss/hidden": 2.9984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29752269983291624,
+      "loss/reg": 0.0,
+      "step": 11880
+    },
+    {
+      "epoch": 0.07822368421052632,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.013792928059895833,
+      "learning_rate": 0.0001,
+      "loss": 3.2052,
+      "loss/crossentropy": 2.5155674695968626,
+      "loss/hidden": 2.9015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2883127599954605,
+      "loss/reg": 0.0,
+      "step": 11890
+    },
+    {
+      "epoch": 0.07828947368421052,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.05334879557291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1696,
+      "loss/crossentropy": 2.3109512329101562,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2651562377810478,
+      "loss/reg": 0.0,
+      "step": 11900
+    },
+    {
+      "epoch": 0.07835526315789473,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.1116363525390625,
+      "learning_rate": 0.0001,
+      "loss": 3.2305,
+      "loss/crossentropy": 2.4196372270584106,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24704778790473939,
+      "loss/reg": 0.0,
+      "step": 11910
+    },
+    {
+      "epoch": 0.07842105263157895,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.10711161295572917,
+      "learning_rate": 0.0001,
+      "loss": 3.2486,
+      "loss/crossentropy": 2.359645998477936,
+      "loss/hidden": 2.9875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28654517233371735,
+      "loss/reg": 0.0,
+      "step": 11920
+    },
+    {
+      "epoch": 0.07848684210526316,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.06504618326822917,
+      "learning_rate": 0.0001,
+      "loss": 3.2312,
+      "loss/crossentropy": 2.364006555080414,
+      "loss/hidden": 2.890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24593105614185334,
+      "loss/reg": 0.0,
+      "step": 11930
+    },
+    {
+      "epoch": 0.07855263157894737,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.11108296712239583,
+      "learning_rate": 0.0001,
+      "loss": 3.1318,
+      "loss/crossentropy": 2.2041036009788515,
+      "loss/hidden": 2.940625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2498387575149536,
+      "loss/reg": 0.0,
+      "step": 11940
+    },
+    {
+      "epoch": 0.07861842105263157,
+      "grad_norm": 3.015625,
+      "grad_norm_var": 0.06920166015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1955,
+      "loss/crossentropy": 2.2502759456634522,
+      "loss/hidden": 3.053125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27016896903514864,
+      "loss/reg": 0.0,
+      "step": 11950
+    },
+    {
+      "epoch": 0.07868421052631579,
+      "grad_norm": 3.125,
+      "grad_norm_var": 0.09563700358072917,
+      "learning_rate": 0.0001,
+      "loss": 3.2128,
+      "loss/crossentropy": 2.1190481543540955,
+      "loss/hidden": 3.1734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.288416750729084,
+      "loss/reg": 0.0,
+      "step": 11960
+    },
+    {
+      "epoch": 0.07875,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.15563863118489582,
+      "learning_rate": 0.0001,
+      "loss": 3.2263,
+      "loss/crossentropy": 1.9541548937559128,
+      "loss/hidden": 2.9234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.271015003323555,
+      "loss/reg": 0.0,
+      "step": 11970
+    },
+    {
+      "epoch": 0.07881578947368421,
+      "grad_norm": 3.015625,
+      "grad_norm_var": 0.09900614420572916,
+      "learning_rate": 0.0001,
+      "loss": 3.0889,
+      "loss/crossentropy": 2.3341493129730226,
+      "loss/hidden": 2.7578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2499636933207512,
+      "loss/reg": 0.0,
+      "step": 11980
+    },
+    {
+      "epoch": 0.07888157894736843,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.07329813639322917,
+      "learning_rate": 0.0001,
+      "loss": 3.1932,
+      "loss/crossentropy": 2.388286221027374,
+      "loss/hidden": 2.85625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24724568724632262,
+      "loss/reg": 0.0,
+      "step": 11990
+    },
+    {
+      "epoch": 0.07894736842105263,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.05429280598958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1876,
+      "loss/crossentropy": 2.351203644275665,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26266307979822157,
+      "loss/reg": 0.0,
+      "step": 12000
+    },
+    {
+      "epoch": 0.07901315789473684,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.04205322265625,
+      "learning_rate": 0.0001,
+      "loss": 3.1379,
+      "loss/crossentropy": 2.3741995811462404,
+      "loss/hidden": 2.8578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24828371405601501,
+      "loss/reg": 0.0,
+      "step": 12010
+    },
+    {
+      "epoch": 0.07907894736842105,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.13201395670572916,
+      "learning_rate": 0.0001,
+      "loss": 3.2741,
+      "loss/crossentropy": 2.037536895275116,
+      "loss/hidden": 2.7109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24230389446020126,
+      "loss/reg": 0.0,
+      "step": 12020
+    },
+    {
+      "epoch": 0.07914473684210527,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.04157613118489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1818,
+      "loss/crossentropy": 2.2516727566719057,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2339026600122452,
+      "loss/reg": 0.0,
+      "step": 12030
+    },
+    {
+      "epoch": 0.07921052631578947,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 26.919873046875,
+      "learning_rate": 0.0001,
+      "loss": 3.2787,
+      "loss/crossentropy": 2.3474916219711304,
+      "loss/hidden": 3.0140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2710026606917381,
+      "loss/reg": 0.0,
+      "step": 12040
+    },
+    {
+      "epoch": 0.07927631578947368,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 26.796516927083335,
+      "learning_rate": 0.0001,
+      "loss": 3.254,
+      "loss/crossentropy": 2.153431460261345,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23484777919948102,
+      "loss/reg": 0.0,
+      "step": 12050
+    },
+    {
+      "epoch": 0.0793421052631579,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.0520416259765625,
+      "learning_rate": 0.0001,
+      "loss": 3.2866,
+      "loss/crossentropy": 2.137704038619995,
+      "loss/hidden": 2.9796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2761687204241753,
+      "loss/reg": 0.0,
+      "step": 12060
+    },
+    {
+      "epoch": 0.07940789473684211,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.05562515258789062,
+      "learning_rate": 0.0001,
+      "loss": 3.1658,
+      "loss/crossentropy": 2.2096517443656922,
+      "loss/hidden": 2.83125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22131142765283585,
+      "loss/reg": 0.0,
+      "step": 12070
+    },
+    {
+      "epoch": 0.07947368421052632,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.033300526936848956,
+      "learning_rate": 0.0001,
+      "loss": 3.1878,
+      "loss/crossentropy": 2.486378014087677,
+      "loss/hidden": 2.965625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.288157556951046,
+      "loss/reg": 0.0,
+      "step": 12080
+    },
+    {
+      "epoch": 0.07953947368421052,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.18038736979166667,
+      "learning_rate": 0.0001,
+      "loss": 3.3441,
+      "loss/crossentropy": 2.1570433020591735,
+      "loss/hidden": 3.3609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33738467693328855,
+      "loss/reg": 0.0,
+      "step": 12090
+    },
+    {
+      "epoch": 0.07960526315789473,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.14531962076822916,
+      "learning_rate": 0.0001,
+      "loss": 3.2092,
+      "loss/crossentropy": 2.11829297542572,
+      "loss/hidden": 3.134375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29526630192995074,
+      "loss/reg": 0.0,
+      "step": 12100
+    },
+    {
+      "epoch": 0.07967105263157895,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.11277669270833333,
+      "learning_rate": 0.0001,
+      "loss": 3.2437,
+      "loss/crossentropy": 2.151422083377838,
+      "loss/hidden": 3.0609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31264509409666064,
+      "loss/reg": 0.0,
+      "step": 12110
+    },
+    {
+      "epoch": 0.07973684210526316,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.09020182291666666,
+      "learning_rate": 0.0001,
+      "loss": 3.1802,
+      "loss/crossentropy": 2.190220355987549,
+      "loss/hidden": 2.9015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26489182114601134,
+      "loss/reg": 0.0,
+      "step": 12120
+    },
+    {
+      "epoch": 0.07980263157894738,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.10075581868489583,
+      "learning_rate": 0.0001,
+      "loss": 3.2733,
+      "loss/crossentropy": 2.4329964399337767,
+      "loss/hidden": 3.0609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29891559183597566,
+      "loss/reg": 0.0,
+      "step": 12130
+    },
+    {
+      "epoch": 0.07986842105263157,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.08171284993489583,
+      "learning_rate": 0.0001,
+      "loss": 3.2143,
+      "loss/crossentropy": 2.3487884759902955,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2699664428830147,
+      "loss/reg": 0.0,
+      "step": 12140
+    },
+    {
+      "epoch": 0.07993421052631579,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.11846415201822917,
+      "learning_rate": 0.0001,
+      "loss": 3.1594,
+      "loss/crossentropy": 2.3031589150428773,
+      "loss/hidden": 2.8625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2605236619710922,
+      "loss/reg": 0.0,
+      "step": 12150
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 3.015625,
+      "grad_norm_var": 0.17751363118489583,
+      "learning_rate": 0.0001,
+      "loss": 3.3283,
+      "loss/crossentropy": 2.4235578894615175,
+      "loss/hidden": 3.265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.40606142282485963,
+      "loss/reg": 0.0,
+      "step": 12160
+    },
+    {
+      "epoch": 0.08006578947368422,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.18465067545572916,
+      "learning_rate": 0.0001,
+      "loss": 3.2069,
+      "loss/crossentropy": 2.3678341031074526,
+      "loss/hidden": 2.93125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26937836706638335,
+      "loss/reg": 0.0,
+      "step": 12170
+    },
+    {
+      "epoch": 0.08013157894736841,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.11225484212239584,
+      "learning_rate": 0.0001,
+      "loss": 3.1606,
+      "loss/crossentropy": 2.606902313232422,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2720926284790039,
+      "loss/reg": 0.0,
+      "step": 12180
+    },
+    {
+      "epoch": 0.08019736842105263,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.03882548014322917,
+      "learning_rate": 0.0001,
+      "loss": 3.2091,
+      "loss/crossentropy": 2.389057195186615,
+      "loss/hidden": 2.934375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2834290415048599,
+      "loss/reg": 0.0,
+      "step": 12190
+    },
+    {
+      "epoch": 0.08026315789473684,
+      "grad_norm": 3.328125,
+      "grad_norm_var": 0.082470703125,
+      "learning_rate": 0.0001,
+      "loss": 3.1846,
+      "loss/crossentropy": 2.1885082483291627,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.250583179295063,
+      "loss/reg": 0.0,
+      "step": 12200
+    },
+    {
+      "epoch": 0.08032894736842106,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.14348551432291667,
+      "learning_rate": 0.0001,
+      "loss": 3.2234,
+      "loss/crossentropy": 2.2288808941841127,
+      "loss/hidden": 3.159375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2857954427599907,
+      "loss/reg": 0.0,
+      "step": 12210
+    },
+    {
+      "epoch": 0.08039473684210527,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.05944010416666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1535,
+      "loss/crossentropy": 2.3943295001983644,
+      "loss/hidden": 2.8984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27064385265111923,
+      "loss/reg": 0.0,
+      "step": 12220
+    },
+    {
+      "epoch": 0.08046052631578947,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.051488240559895836,
+      "learning_rate": 0.0001,
+      "loss": 3.2165,
+      "loss/crossentropy": 2.36237952709198,
+      "loss/hidden": 3.1921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29628041982650755,
+      "loss/reg": 0.0,
+      "step": 12230
+    },
+    {
+      "epoch": 0.08052631578947368,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.06207682291666667,
+      "learning_rate": 0.0001,
+      "loss": 3.2112,
+      "loss/crossentropy": 2.26582453250885,
+      "loss/hidden": 2.95625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3621983379125595,
+      "loss/reg": 0.0,
+      "step": 12240
+    },
+    {
+      "epoch": 0.0805921052631579,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.08144124348958333,
+      "learning_rate": 0.0001,
+      "loss": 3.2711,
+      "loss/crossentropy": 2.4002971291542052,
+      "loss/hidden": 3.00625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29988196343183515,
+      "loss/reg": 0.0,
+      "step": 12250
+    },
+    {
+      "epoch": 0.08065789473684211,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 3.1181549072265624,
+      "learning_rate": 0.0001,
+      "loss": 3.2486,
+      "loss/crossentropy": 2.26135613322258,
+      "loss/hidden": 2.8453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24128984957933425,
+      "loss/reg": 0.0,
+      "step": 12260
+    },
+    {
+      "epoch": 0.08072368421052632,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 4.51636962890625,
+      "learning_rate": 0.0001,
+      "loss": 3.5227,
+      "loss/crossentropy": 2.2277899503707888,
+      "loss/hidden": 3.453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35484138429164885,
+      "loss/reg": 0.0,
+      "step": 12270
+    },
+    {
+      "epoch": 0.08078947368421052,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 2.06083984375,
+      "learning_rate": 0.0001,
+      "loss": 3.2936,
+      "loss/crossentropy": 2.313612127304077,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2666293315589428,
+      "loss/reg": 0.0,
+      "step": 12280
+    },
+    {
+      "epoch": 0.08085526315789474,
+      "grad_norm": 4.625,
+      "grad_norm_var": 6.363402303059896,
+      "learning_rate": 0.0001,
+      "loss": 3.2228,
+      "loss/crossentropy": 2.3592599511146544,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25869676619768145,
+      "loss/reg": 0.0,
+      "step": 12290
+    },
+    {
+      "epoch": 0.08092105263157895,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.3772532145182292,
+      "learning_rate": 0.0001,
+      "loss": 3.2186,
+      "loss/crossentropy": 2.50057338476181,
+      "loss/hidden": 2.9875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2711464300751686,
+      "loss/reg": 0.0,
+      "step": 12300
+    },
+    {
+      "epoch": 0.08098684210526316,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.07467447916666667,
+      "learning_rate": 0.0001,
+      "loss": 3.2317,
+      "loss/crossentropy": 2.381076216697693,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24998296648263932,
+      "loss/reg": 0.0,
+      "step": 12310
+    },
+    {
+      "epoch": 0.08105263157894736,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 1.58385009765625,
+      "learning_rate": 0.0001,
+      "loss": 3.3666,
+      "loss/crossentropy": 2.2987404227256776,
+      "loss/hidden": 2.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24597887545824051,
+      "loss/reg": 0.0,
+      "step": 12320
+    },
+    {
+      "epoch": 0.08111842105263158,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 3.158852918199495e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.3716,
+      "loss/crossentropy": 2.415051448345184,
+      "loss/hidden": 2.9515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29122400283813477,
+      "loss/reg": 0.0,
+      "step": 12330
+    },
+    {
+      "epoch": 0.08118421052631579,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 3.158852918454168e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2774,
+      "loss/crossentropy": 2.082320672273636,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2748811081051826,
+      "loss/reg": 0.0,
+      "step": 12340
+    },
+    {
+      "epoch": 0.08125,
+      "grad_norm": 3.46875,
+      "grad_norm_var": 0.26638997395833336,
+      "learning_rate": 0.0001,
+      "loss": 3.2315,
+      "loss/crossentropy": 2.7044607162475587,
+      "loss/hidden": 2.953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32181375473737717,
+      "loss/reg": 0.0,
+      "step": 12350
+    },
+    {
+      "epoch": 0.08131578947368422,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.22693684895833333,
+      "learning_rate": 0.0001,
+      "loss": 3.2351,
+      "loss/crossentropy": 2.7221840620040894,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26635742783546446,
+      "loss/reg": 0.0,
+      "step": 12360
+    },
+    {
+      "epoch": 0.08138157894736842,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.26437174479166664,
+      "learning_rate": 0.0001,
+      "loss": 3.2068,
+      "loss/crossentropy": 2.3512622594833372,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2477712720632553,
+      "loss/reg": 0.0,
+      "step": 12370
+    },
+    {
+      "epoch": 0.08144736842105263,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.3268707275390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1974,
+      "loss/crossentropy": 2.210025131702423,
+      "loss/hidden": 3.05625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36077398508787156,
+      "loss/reg": 0.0,
+      "step": 12380
+    },
+    {
+      "epoch": 0.08151315789473684,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.5682525634765625,
+      "learning_rate": 0.0001,
+      "loss": 3.2079,
+      "loss/crossentropy": 2.0402897000312805,
+      "loss/hidden": 2.965625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26192123591899874,
+      "loss/reg": 0.0,
+      "step": 12390
+    },
+    {
+      "epoch": 0.08157894736842106,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.3712565104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.2107,
+      "loss/crossentropy": 1.9560218453407288,
+      "loss/hidden": 2.925,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21844983994960784,
+      "loss/reg": 0.0,
+      "step": 12400
+    },
+    {
+      "epoch": 0.08164473684210527,
+      "grad_norm": 3.171875,
+      "grad_norm_var": 0.4505859375,
+      "learning_rate": 0.0001,
+      "loss": 3.2759,
+      "loss/crossentropy": 2.3604748249053955,
+      "loss/hidden": 2.85625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2524840489029884,
+      "loss/reg": 0.0,
+      "step": 12410
+    },
+    {
+      "epoch": 0.08171052631578947,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.24068603515625,
+      "learning_rate": 0.0001,
+      "loss": 3.2728,
+      "loss/crossentropy": 2.517817199230194,
+      "loss/hidden": 3.0015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29347013384103776,
+      "loss/reg": 0.0,
+      "step": 12420
+    },
+    {
+      "epoch": 0.08177631578947368,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.12271728515625,
+      "learning_rate": 0.0001,
+      "loss": 3.2551,
+      "loss/crossentropy": 2.4665472149848937,
+      "loss/hidden": 2.93125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2709146931767464,
+      "loss/reg": 0.0,
+      "step": 12430
+    },
+    {
+      "epoch": 0.0818421052631579,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.2295074462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.2219,
+      "loss/crossentropy": 2.40498046875,
+      "loss/hidden": 3.1,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32999152690172195,
+      "loss/reg": 0.0,
+      "step": 12440
+    },
+    {
+      "epoch": 0.08190789473684211,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.1619049072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0867,
+      "loss/crossentropy": 2.3188422203063963,
+      "loss/hidden": 2.8984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29690912514925005,
+      "loss/reg": 0.0,
+      "step": 12450
+    },
+    {
+      "epoch": 0.08197368421052631,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.07049051920572917,
+      "learning_rate": 0.0001,
+      "loss": 3.2682,
+      "loss/crossentropy": 2.588274967670441,
+      "loss/hidden": 3.4046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.39090928733348845,
+      "loss/reg": 0.0,
+      "step": 12460
+    },
+    {
+      "epoch": 0.08203947368421052,
+      "grad_norm": 3.921875,
+      "grad_norm_var": 0.2250261942545573,
+      "learning_rate": 0.0001,
+      "loss": 3.2503,
+      "loss/crossentropy": 2.6322230458259583,
+      "loss/hidden": 2.828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2629552409052849,
+      "loss/reg": 0.0,
+      "step": 12470
+    },
+    {
+      "epoch": 0.08210526315789474,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.18522109985351562,
+      "learning_rate": 0.0001,
+      "loss": 3.1313,
+      "loss/crossentropy": 2.1257114171981812,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23226768374443055,
+      "loss/reg": 0.0,
+      "step": 12480
+    },
+    {
+      "epoch": 0.08217105263157895,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.16383056640625,
+      "learning_rate": 0.0001,
+      "loss": 3.2008,
+      "loss/crossentropy": 1.9993813276290893,
+      "loss/hidden": 2.821875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21686773076653482,
+      "loss/reg": 0.0,
+      "step": 12490
+    },
+    {
+      "epoch": 0.08223684210526316,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.12244364420572916,
+      "learning_rate": 0.0001,
+      "loss": 3.1643,
+      "loss/crossentropy": 2.4217318654060365,
+      "loss/hidden": 2.8796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2706719309091568,
+      "loss/reg": 0.0,
+      "step": 12500
+    },
+    {
+      "epoch": 0.08230263157894736,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.9795888264973959,
+      "learning_rate": 0.0001,
+      "loss": 3.2763,
+      "loss/crossentropy": 2.131495940685272,
+      "loss/hidden": 2.9625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24229931831359863,
+      "loss/reg": 0.0,
+      "step": 12510
+    },
+    {
+      "epoch": 0.08236842105263158,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 1.021240234375,
+      "learning_rate": 0.0001,
+      "loss": 3.1894,
+      "loss/crossentropy": 2.2545747995376586,
+      "loss/hidden": 2.9140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26802987307310105,
+      "loss/reg": 0.0,
+      "step": 12520
+    },
+    {
+      "epoch": 0.08243421052631579,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 2.0329969940865024e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.364,
+      "loss/crossentropy": 2.37786750793457,
+      "loss/hidden": 2.9140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27177259773015977,
+      "loss/reg": 0.0,
+      "step": 12530
+    },
+    {
+      "epoch": 0.0825,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.0586822509765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1425,
+      "loss/crossentropy": 2.3742210388183596,
+      "loss/hidden": 2.9078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25972897857427596,
+      "loss/reg": 0.0,
+      "step": 12540
+    },
+    {
+      "epoch": 0.08256578947368422,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.16419270833333333,
+      "learning_rate": 0.0001,
+      "loss": 3.1808,
+      "loss/crossentropy": 2.2249507308006287,
+      "loss/hidden": 2.9671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.260022896528244,
+      "loss/reg": 0.0,
+      "step": 12550
+    },
+    {
+      "epoch": 0.08263157894736842,
+      "grad_norm": 3.765625,
+      "grad_norm_var": 0.1635406494140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2244,
+      "loss/crossentropy": 2.417391860485077,
+      "loss/hidden": 2.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24499612003564836,
+      "loss/reg": 0.0,
+      "step": 12560
+    },
+    {
+      "epoch": 0.08269736842105263,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.24940999348958334,
+      "learning_rate": 0.0001,
+      "loss": 3.259,
+      "loss/crossentropy": 2.008757221698761,
+      "loss/hidden": 3.371875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28915109634399416,
+      "loss/reg": 0.0,
+      "step": 12570
+    },
+    {
+      "epoch": 0.08276315789473684,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.030516560872395834,
+      "learning_rate": 0.0001,
+      "loss": 3.0918,
+      "loss/crossentropy": 2.3331239223480225,
+      "loss/hidden": 2.840625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27426364421844485,
+      "loss/reg": 0.0,
+      "step": 12580
+    },
+    {
+      "epoch": 0.08282894736842106,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.4984527587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.2675,
+      "loss/crossentropy": 2.3463852405548096,
+      "loss/hidden": 3.0046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2746781826019287,
+      "loss/reg": 0.0,
+      "step": 12590
+    },
+    {
+      "epoch": 0.08289473684210526,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.026822916666666665,
+      "learning_rate": 0.0001,
+      "loss": 3.0942,
+      "loss/crossentropy": 2.1946144729852675,
+      "loss/hidden": 3.0390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.273574560880661,
+      "loss/reg": 0.0,
+      "step": 12600
+    },
+    {
+      "epoch": 0.08296052631578947,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.399658203125,
+      "learning_rate": 0.0001,
+      "loss": 3.3113,
+      "loss/crossentropy": 2.292886030673981,
+      "loss/hidden": 3.0203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3399674043059349,
+      "loss/reg": 0.0,
+      "step": 12610
+    },
+    {
+      "epoch": 0.08302631578947368,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.306982421875,
+      "learning_rate": 0.0001,
+      "loss": 3.3316,
+      "loss/crossentropy": 2.2384460091590883,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2688921958208084,
+      "loss/reg": 0.0,
+      "step": 12620
+    },
+    {
+      "epoch": 0.0830921052631579,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.04449462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1944,
+      "loss/crossentropy": 2.2317716479301453,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23216352015733718,
+      "loss/reg": 0.0,
+      "step": 12630
+    },
+    {
+      "epoch": 0.08315789473684211,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.05810139973958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1734,
+      "loss/crossentropy": 2.513770651817322,
+      "loss/hidden": 2.840625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26424300968647,
+      "loss/reg": 0.0,
+      "step": 12640
+    },
+    {
+      "epoch": 0.08322368421052631,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.060302734375,
+      "learning_rate": 0.0001,
+      "loss": 3.212,
+      "loss/crossentropy": 2.3643002271652223,
+      "loss/hidden": 2.9875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2974396377801895,
+      "loss/reg": 0.0,
+      "step": 12650
+    },
+    {
+      "epoch": 0.08328947368421052,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.11692301432291667,
+      "learning_rate": 0.0001,
+      "loss": 3.2013,
+      "loss/crossentropy": 2.200558376312256,
+      "loss/hidden": 2.934375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26023727655410767,
+      "loss/reg": 0.0,
+      "step": 12660
+    },
+    {
+      "epoch": 0.08335526315789474,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.0976470947265625,
+      "learning_rate": 0.0001,
+      "loss": 3.2425,
+      "loss/crossentropy": 2.3456878662109375,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30842293947935107,
+      "loss/reg": 0.0,
+      "step": 12670
+    },
+    {
+      "epoch": 0.08342105263157895,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.05621744791666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1823,
+      "loss/crossentropy": 2.2649134039878844,
+      "loss/hidden": 2.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24366314560174943,
+      "loss/reg": 0.0,
+      "step": 12680
+    },
+    {
+      "epoch": 0.08348684210526315,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.059891764322916666,
+      "learning_rate": 0.0001,
+      "loss": 3.1859,
+      "loss/crossentropy": 2.0741775274276733,
+      "loss/hidden": 2.946875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24329135864973067,
+      "loss/reg": 0.0,
+      "step": 12690
+    },
+    {
+      "epoch": 0.08355263157894736,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.13880208333333333,
+      "learning_rate": 0.0001,
+      "loss": 3.1551,
+      "loss/crossentropy": 2.4719223856925963,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24188547879457473,
+      "loss/reg": 0.0,
+      "step": 12700
+    },
+    {
+      "epoch": 0.08361842105263158,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.32080078125,
+      "learning_rate": 0.0001,
+      "loss": 3.3068,
+      "loss/crossentropy": 2.279301416873932,
+      "loss/hidden": 2.95625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2998178914189339,
+      "loss/reg": 0.0,
+      "step": 12710
+    },
+    {
+      "epoch": 0.08368421052631579,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.341064453125,
+      "learning_rate": 0.0001,
+      "loss": 3.2089,
+      "loss/crossentropy": 2.4134485125541687,
+      "loss/hidden": 2.9875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28444976508617403,
+      "loss/reg": 0.0,
+      "step": 12720
+    },
+    {
+      "epoch": 0.08375,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.13637593587239583,
+      "learning_rate": 0.0001,
+      "loss": 3.2345,
+      "loss/crossentropy": 2.4403869032859804,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26378336995840074,
+      "loss/reg": 0.0,
+      "step": 12730
+    },
+    {
+      "epoch": 0.0838157894736842,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.0974273681640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1501,
+      "loss/crossentropy": 2.1977667093276976,
+      "loss/hidden": 2.9296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23563524186611176,
+      "loss/reg": 0.0,
+      "step": 12740
+    },
+    {
+      "epoch": 0.08388157894736842,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.12566731770833334,
+      "learning_rate": 0.0001,
+      "loss": 3.2516,
+      "loss/crossentropy": 2.206720507144928,
+      "loss/hidden": 3.0453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2731472015380859,
+      "loss/reg": 0.0,
+      "step": 12750
+    },
+    {
+      "epoch": 0.08394736842105263,
+      "grad_norm": 3.859375,
+      "grad_norm_var": 0.21768290201822918,
+      "learning_rate": 0.0001,
+      "loss": 3.2035,
+      "loss/crossentropy": 2.3951833486557006,
+      "loss/hidden": 2.8984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26070789247751236,
+      "loss/reg": 0.0,
+      "step": 12760
+    },
+    {
+      "epoch": 0.08401315789473685,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.36861572265625,
+      "learning_rate": 0.0001,
+      "loss": 3.3054,
+      "loss/crossentropy": 2.393438732624054,
+      "loss/hidden": 3.2328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3618380635976791,
+      "loss/reg": 0.0,
+      "step": 12770
+    },
+    {
+      "epoch": 0.08407894736842106,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.24163004557291667,
+      "learning_rate": 0.0001,
+      "loss": 3.2972,
+      "loss/crossentropy": 2.510524129867554,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26494802087545394,
+      "loss/reg": 0.0,
+      "step": 12780
+    },
+    {
+      "epoch": 0.08414473684210526,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.07789713541666667,
+      "learning_rate": 0.0001,
+      "loss": 3.2375,
+      "loss/crossentropy": 2.326508915424347,
+      "loss/hidden": 2.8578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25067940801382066,
+      "loss/reg": 0.0,
+      "step": 12790
+    },
+    {
+      "epoch": 0.08421052631578947,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.0654693603515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1664,
+      "loss/crossentropy": 2.2524615049362184,
+      "loss/hidden": 2.9640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23585905730724335,
+      "loss/reg": 0.0,
+      "step": 12800
+    },
+    {
+      "epoch": 0.08427631578947369,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.05461324055989583,
+      "learning_rate": 0.0001,
+      "loss": 3.2086,
+      "loss/crossentropy": 2.586345672607422,
+      "loss/hidden": 2.9109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2775671869516373,
+      "loss/reg": 0.0,
+      "step": 12810
+    },
+    {
+      "epoch": 0.0843421052631579,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.157568359375,
+      "learning_rate": 0.0001,
+      "loss": 3.2467,
+      "loss/crossentropy": 2.4124717354774474,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27734949439764023,
+      "loss/reg": 0.0,
+      "step": 12820
+    },
+    {
+      "epoch": 0.0844078947368421,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.151953125,
+      "learning_rate": 0.0001,
+      "loss": 3.1838,
+      "loss/crossentropy": 2.217060422897339,
+      "loss/hidden": 2.940625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31069841980934143,
+      "loss/reg": 0.0,
+      "step": 12830
+    },
+    {
+      "epoch": 0.08447368421052631,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.15916341145833332,
+      "learning_rate": 0.0001,
+      "loss": 3.251,
+      "loss/crossentropy": 2.2915706515312193,
+      "loss/hidden": 3.128125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28919376283884046,
+      "loss/reg": 0.0,
+      "step": 12840
+    },
+    {
+      "epoch": 0.08453947368421053,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.05465494791666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1509,
+      "loss/crossentropy": 2.3436198830604553,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2566779345273972,
+      "loss/reg": 0.0,
+      "step": 12850
+    },
+    {
+      "epoch": 0.08460526315789474,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.12815653483072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1553,
+      "loss/crossentropy": 2.6007506489753722,
+      "loss/hidden": 2.91875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27069382518529894,
+      "loss/reg": 0.0,
+      "step": 12860
+    },
+    {
+      "epoch": 0.08467105263157895,
+      "grad_norm": 3.125,
+      "grad_norm_var": 0.09325764973958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1848,
+      "loss/crossentropy": 2.3489827513694763,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2542911395430565,
+      "loss/reg": 0.0,
+      "step": 12870
+    },
+    {
+      "epoch": 0.08473684210526315,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.2053375244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2528,
+      "loss/crossentropy": 2.3320749402046204,
+      "loss/hidden": 2.9296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25249498784542085,
+      "loss/reg": 0.0,
+      "step": 12880
+    },
+    {
+      "epoch": 0.08480263157894737,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.22139383951822916,
+      "learning_rate": 0.0001,
+      "loss": 3.2744,
+      "loss/crossentropy": 2.5565970659255983,
+      "loss/hidden": 3.021875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30419613122940065,
+      "loss/reg": 0.0,
+      "step": 12890
+    },
+    {
+      "epoch": 0.08486842105263158,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.12555338541666666,
+      "learning_rate": 0.0001,
+      "loss": 3.2548,
+      "loss/crossentropy": 2.382706320285797,
+      "loss/hidden": 3.0875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28827311396598815,
+      "loss/reg": 0.0,
+      "step": 12900
+    },
+    {
+      "epoch": 0.08493421052631579,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.031119791666666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1646,
+      "loss/crossentropy": 2.209449625015259,
+      "loss/hidden": 3.0421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27490794360637666,
+      "loss/reg": 0.0,
+      "step": 12910
+    },
+    {
+      "epoch": 0.085,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.075537109375,
+      "learning_rate": 0.0001,
+      "loss": 3.1151,
+      "loss/crossentropy": 2.42991498708725,
+      "loss/hidden": 2.9171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24879284277558328,
+      "loss/reg": 0.0,
+      "step": 12920
+    },
+    {
+      "epoch": 0.0850657894736842,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.06396865844726562,
+      "learning_rate": 0.0001,
+      "loss": 3.1176,
+      "loss/crossentropy": 2.124844658374786,
+      "loss/hidden": 2.9796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26795649230480195,
+      "loss/reg": 0.0,
+      "step": 12930
+    },
+    {
+      "epoch": 0.08513157894736842,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.0884844462076823,
+      "learning_rate": 0.0001,
+      "loss": 3.1416,
+      "loss/crossentropy": 2.390605056285858,
+      "loss/hidden": 2.8765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2603973612189293,
+      "loss/reg": 0.0,
+      "step": 12940
+    },
+    {
+      "epoch": 0.08519736842105263,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.11295572916666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1506,
+      "loss/crossentropy": 2.3307228684425354,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23968299478292465,
+      "loss/reg": 0.0,
+      "step": 12950
+    },
+    {
+      "epoch": 0.08526315789473685,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.07505594889322917,
+      "learning_rate": 0.0001,
+      "loss": 3.1409,
+      "loss/crossentropy": 2.32118815779686,
+      "loss/hidden": 2.965625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27437605410814286,
+      "loss/reg": 0.0,
+      "step": 12960
+    },
+    {
+      "epoch": 0.08532894736842105,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.05011393229166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1943,
+      "loss/crossentropy": 2.1953859329223633,
+      "loss/hidden": 2.95625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2964278385043144,
+      "loss/reg": 0.0,
+      "step": 12970
+    },
+    {
+      "epoch": 0.08539473684210526,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.044733683268229164,
+      "learning_rate": 0.0001,
+      "loss": 3.1596,
+      "loss/crossentropy": 2.48675742149353,
+      "loss/hidden": 2.95,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2866469621658325,
+      "loss/reg": 0.0,
+      "step": 12980
+    },
+    {
+      "epoch": 0.08546052631578947,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.04439188639322917,
+      "learning_rate": 0.0001,
+      "loss": 3.1955,
+      "loss/crossentropy": 2.3276284098625184,
+      "loss/hidden": 2.78125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.254511134326458,
+      "loss/reg": 0.0,
+      "step": 12990
+    },
+    {
+      "epoch": 0.08552631578947369,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.18092041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.3308,
+      "loss/crossentropy": 2.436275231838226,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27356108725070954,
+      "loss/reg": 0.0,
+      "step": 13000
+    },
+    {
+      "epoch": 0.0855921052631579,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.17511393229166666,
+      "learning_rate": 0.0001,
+      "loss": 3.1302,
+      "loss/crossentropy": 2.2856626510620117,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24442512094974517,
+      "loss/reg": 0.0,
+      "step": 13010
+    },
+    {
+      "epoch": 0.0856578947368421,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.10161844889322917,
+      "learning_rate": 0.0001,
+      "loss": 3.1923,
+      "loss/crossentropy": 2.4494728326797484,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2575364217162132,
+      "loss/reg": 0.0,
+      "step": 13020
+    },
+    {
+      "epoch": 0.08572368421052631,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.09909566243489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1732,
+      "loss/crossentropy": 2.4833264112472535,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2727135464549065,
+      "loss/reg": 0.0,
+      "step": 13030
+    },
+    {
+      "epoch": 0.08578947368421053,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.12711588541666666,
+      "learning_rate": 0.0001,
+      "loss": 3.165,
+      "loss/crossentropy": 2.284275805950165,
+      "loss/hidden": 3.0859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29309146106243134,
+      "loss/reg": 0.0,
+      "step": 13040
+    },
+    {
+      "epoch": 0.08585526315789474,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.18906148274739584,
+      "learning_rate": 0.0001,
+      "loss": 3.1988,
+      "loss/crossentropy": 2.1231042385101317,
+      "loss/hidden": 3.0375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.246621835231781,
+      "loss/reg": 0.0,
+      "step": 13050
+    },
+    {
+      "epoch": 0.08592105263157895,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.14317118326822917,
+      "learning_rate": 0.0001,
+      "loss": 3.2204,
+      "loss/crossentropy": 2.115370142459869,
+      "loss/hidden": 2.7125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22535212635993956,
+      "loss/reg": 0.0,
+      "step": 13060
+    },
+    {
+      "epoch": 0.08598684210526315,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.17832743326822917,
+      "learning_rate": 0.0001,
+      "loss": 3.1988,
+      "loss/crossentropy": 2.271882343292236,
+      "loss/hidden": 3.0140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2518279105424881,
+      "loss/reg": 0.0,
+      "step": 13070
+    },
+    {
+      "epoch": 0.08605263157894737,
+      "grad_norm": 1.90625,
+      "grad_norm_var": 0.1100250244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2672,
+      "loss/crossentropy": 2.211618059873581,
+      "loss/hidden": 3.015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2709794193506241,
+      "loss/reg": 0.0,
+      "step": 13080
+    },
+    {
+      "epoch": 0.08611842105263158,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.06676432291666666,
+      "learning_rate": 0.0001,
+      "loss": 3.2393,
+      "loss/crossentropy": 2.4678762197494506,
+      "loss/hidden": 2.9453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2634593158960342,
+      "loss/reg": 0.0,
+      "step": 13090
+    },
+    {
+      "epoch": 0.0861842105263158,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.030402628580729167,
+      "learning_rate": 0.0001,
+      "loss": 3.19,
+      "loss/crossentropy": 2.5426036715507507,
+      "loss/hidden": 2.9296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26041875034570694,
+      "loss/reg": 0.0,
+      "step": 13100
+    },
+    {
+      "epoch": 0.08625,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.028034464518229166,
+      "learning_rate": 0.0001,
+      "loss": 3.1556,
+      "loss/crossentropy": 2.1748136937618257,
+      "loss/hidden": 3.0421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3230336934328079,
+      "loss/reg": 0.0,
+      "step": 13110
+    },
+    {
+      "epoch": 0.0863157894736842,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.3787750244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2653,
+      "loss/crossentropy": 2.352058470249176,
+      "loss/hidden": 2.89375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3100366100668907,
+      "loss/reg": 0.0,
+      "step": 13120
+    },
+    {
+      "epoch": 0.08638157894736842,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.12065327962239583,
+      "learning_rate": 0.0001,
+      "loss": 3.2897,
+      "loss/crossentropy": 2.2920926332473757,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24493622779846191,
+      "loss/reg": 0.0,
+      "step": 13130
+    },
+    {
+      "epoch": 0.08644736842105263,
+      "grad_norm": 3.140625,
+      "grad_norm_var": 0.17266337076822916,
+      "learning_rate": 0.0001,
+      "loss": 3.2541,
+      "loss/crossentropy": 2.253483748435974,
+      "loss/hidden": 2.815625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24110897034406661,
+      "loss/reg": 0.0,
+      "step": 13140
+    },
+    {
+      "epoch": 0.08651315789473685,
+      "grad_norm": 3.140625,
+      "grad_norm_var": 0.15883687337239583,
+      "learning_rate": 0.0001,
+      "loss": 3.2161,
+      "loss/crossentropy": 2.490678381919861,
+      "loss/hidden": 2.9890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2905942976474762,
+      "loss/reg": 0.0,
+      "step": 13150
+    },
+    {
+      "epoch": 0.08657894736842105,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.124365234375,
+      "learning_rate": 0.0001,
+      "loss": 3.2037,
+      "loss/crossentropy": 2.607399010658264,
+      "loss/hidden": 3.0484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3426252081990242,
+      "loss/reg": 0.0,
+      "step": 13160
+    },
+    {
+      "epoch": 0.08664473684210526,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.20832697550455728,
+      "learning_rate": 0.0001,
+      "loss": 3.3389,
+      "loss/crossentropy": 2.5389102935791015,
+      "loss/hidden": 2.90625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27587536424398423,
+      "loss/reg": 0.0,
+      "step": 13170
+    },
+    {
+      "epoch": 0.08671052631578947,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 3.749950368844328e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.3435,
+      "loss/crossentropy": 2.2043145060539246,
+      "loss/hidden": 3.0921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27380194514989853,
+      "loss/reg": 0.0,
+      "step": 13180
+    },
+    {
+      "epoch": 0.08677631578947369,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.10250651041666667,
+      "learning_rate": 0.0001,
+      "loss": 3.2179,
+      "loss/crossentropy": 2.5562551975250245,
+      "loss/hidden": 3.125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35649020969867706,
+      "loss/reg": 0.0,
+      "step": 13190
+    },
+    {
+      "epoch": 0.0868421052631579,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.11367899576822917,
+      "learning_rate": 0.0001,
+      "loss": 3.1639,
+      "loss/crossentropy": 2.2194557189941406,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2623781323432922,
+      "loss/reg": 0.0,
+      "step": 13200
+    },
+    {
+      "epoch": 0.0869078947368421,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.09927978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.203,
+      "loss/crossentropy": 2.3938650250434876,
+      "loss/hidden": 2.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22924545854330064,
+      "loss/reg": 0.0,
+      "step": 13210
+    },
+    {
+      "epoch": 0.08697368421052631,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.07858072916666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1885,
+      "loss/crossentropy": 2.659112477302551,
+      "loss/hidden": 3.2625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3096113160252571,
+      "loss/reg": 0.0,
+      "step": 13220
+    },
+    {
+      "epoch": 0.08703947368421053,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.08105061848958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1608,
+      "loss/crossentropy": 2.2919702410697935,
+      "loss/hidden": 3.0671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28502654284238815,
+      "loss/reg": 0.0,
+      "step": 13230
+    },
+    {
+      "epoch": 0.08710526315789474,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.2513631184895833,
+      "learning_rate": 0.0001,
+      "loss": 3.1995,
+      "loss/crossentropy": 2.206997013092041,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20357893258333207,
+      "loss/reg": 0.0,
+      "step": 13240
+    },
+    {
+      "epoch": 0.08717105263157894,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.06531575520833334,
+      "learning_rate": 0.0001,
+      "loss": 3.1267,
+      "loss/crossentropy": 2.095241755247116,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23652465790510177,
+      "loss/reg": 0.0,
+      "step": 13250
+    },
+    {
+      "epoch": 0.08723684210526315,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.110009765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1799,
+      "loss/crossentropy": 2.1254674077033995,
+      "loss/hidden": 2.6765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22684407681226731,
+      "loss/reg": 0.0,
+      "step": 13260
+    },
+    {
+      "epoch": 0.08730263157894737,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 3.398986257643471e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.358,
+      "loss/crossentropy": 2.3295519828796385,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30277568846940994,
+      "loss/reg": 0.0,
+      "step": 13270
+    },
+    {
+      "epoch": 0.08736842105263158,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 3.3989862579380115e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.3036,
+      "loss/crossentropy": 2.496494376659393,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2583273336291313,
+      "loss/reg": 0.0,
+      "step": 13280
+    },
+    {
+      "epoch": 0.0874342105263158,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.0663238525390625,
+      "learning_rate": 0.0001,
+      "loss": 3.2025,
+      "loss/crossentropy": 2.332583689689636,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24232984483242034,
+      "loss/reg": 0.0,
+      "step": 13290
+    },
+    {
+      "epoch": 0.0875,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.43166910807291664,
+      "learning_rate": 0.0001,
+      "loss": 3.1339,
+      "loss/crossentropy": 2.394269013404846,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24607907682657243,
+      "loss/reg": 0.0,
+      "step": 13300
+    },
+    {
+      "epoch": 0.08756578947368421,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.112890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1815,
+      "loss/crossentropy": 2.2034417927265166,
+      "loss/hidden": 3.0546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25049934834241866,
+      "loss/reg": 0.0,
+      "step": 13310
+    },
+    {
+      "epoch": 0.08763157894736842,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.0748443603515625,
+      "learning_rate": 0.0001,
+      "loss": 3.2491,
+      "loss/crossentropy": 2.058911919593811,
+      "loss/hidden": 3.1359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25788910537958143,
+      "loss/reg": 0.0,
+      "step": 13320
+    },
+    {
+      "epoch": 0.08769736842105263,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.22857666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.2051,
+      "loss/crossentropy": 2.3308457016944883,
+      "loss/hidden": 2.91875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3006305813789368,
+      "loss/reg": 0.0,
+      "step": 13330
+    },
+    {
+      "epoch": 0.08776315789473685,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.266796875,
+      "learning_rate": 0.0001,
+      "loss": 3.1792,
+      "loss/crossentropy": 2.392533528804779,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25690300911664965,
+      "loss/reg": 0.0,
+      "step": 13340
+    },
+    {
+      "epoch": 0.08782894736842105,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.24251302083333334,
+      "learning_rate": 0.0001,
+      "loss": 3.1342,
+      "loss/crossentropy": 2.303742027282715,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24467559903860092,
+      "loss/reg": 0.0,
+      "step": 13350
+    },
+    {
+      "epoch": 0.08789473684210526,
+      "grad_norm": 3.21875,
+      "grad_norm_var": 0.26910400390625,
+      "learning_rate": 0.0001,
+      "loss": 3.2614,
+      "loss/crossentropy": 2.5700215101242065,
+      "loss/hidden": 3.0078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.346772675216198,
+      "loss/reg": 0.0,
+      "step": 13360
+    },
+    {
+      "epoch": 0.08796052631578948,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 8.30354715983073,
+      "learning_rate": 0.0001,
+      "loss": 3.1987,
+      "loss/crossentropy": 2.3953630328178406,
+      "loss/hidden": 2.884375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2601369693875313,
+      "loss/reg": 0.0,
+      "step": 13370
+    },
+    {
+      "epoch": 0.08802631578947369,
+      "grad_norm": 3.671875,
+      "grad_norm_var": 3.5722076416015627,
+      "learning_rate": 0.0001,
+      "loss": 3.2285,
+      "loss/crossentropy": 2.540582847595215,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.274001345038414,
+      "loss/reg": 0.0,
+      "step": 13380
+    },
+    {
+      "epoch": 0.08809210526315789,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.2377593994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1787,
+      "loss/crossentropy": 2.3221506476402283,
+      "loss/hidden": 2.86875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2874672919511795,
+      "loss/reg": 0.0,
+      "step": 13390
+    },
+    {
+      "epoch": 0.0881578947368421,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 1.4219309488932292,
+      "learning_rate": 0.0001,
+      "loss": 3.1862,
+      "loss/crossentropy": 2.255697971582413,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25251765847206115,
+      "loss/reg": 0.0,
+      "step": 13400
+    },
+    {
+      "epoch": 0.08822368421052632,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.023307291666666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1642,
+      "loss/crossentropy": 2.376703941822052,
+      "loss/hidden": 2.8578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2795195817947388,
+      "loss/reg": 0.0,
+      "step": 13410
+    },
+    {
+      "epoch": 0.08828947368421053,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.32665608723958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1802,
+      "loss/crossentropy": 2.390151119232178,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.255537736415863,
+      "loss/reg": 0.0,
+      "step": 13420
+    },
+    {
+      "epoch": 0.08835526315789474,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.33414306640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1779,
+      "loss/crossentropy": 2.2273300528526305,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2409852236509323,
+      "loss/reg": 0.0,
+      "step": 13430
+    },
+    {
+      "epoch": 0.08842105263157894,
+      "grad_norm": 25.75,
+      "grad_norm_var": 34.02027587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1974,
+      "loss/crossentropy": 2.281948208808899,
+      "loss/hidden": 2.7515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2378145158290863,
+      "loss/reg": 0.0,
+      "step": 13440
+    },
+    {
+      "epoch": 0.08848684210526316,
+      "grad_norm": 2.75,
+      "grad_norm_var": 34.235252888997394,
+      "learning_rate": 0.0001,
+      "loss": 3.3736,
+      "loss/crossentropy": 2.267159104347229,
+      "loss/hidden": 3.1296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2702139914035797,
+      "loss/reg": 0.0,
+      "step": 13450
+    },
+    {
+      "epoch": 0.08855263157894737,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 1.1851145426432292,
+      "learning_rate": 0.0001,
+      "loss": 3.16,
+      "loss/crossentropy": 2.3185499548912047,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2232184700667858,
+      "loss/reg": 0.0,
+      "step": 13460
+    },
+    {
+      "epoch": 0.08861842105263158,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.07108968098958333,
+      "learning_rate": 0.0001,
+      "loss": 3.2029,
+      "loss/crossentropy": 2.1752323627471926,
+      "loss/hidden": 2.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3101599723100662,
+      "loss/reg": 0.0,
+      "step": 13470
+    },
+    {
+      "epoch": 0.0886842105263158,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.11365559895833334,
+      "learning_rate": 0.0001,
+      "loss": 3.2671,
+      "loss/crossentropy": 2.255975532531738,
+      "loss/hidden": 3.125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30118285566568376,
+      "loss/reg": 0.0,
+      "step": 13480
+    },
+    {
+      "epoch": 0.08875,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.06363525390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1907,
+      "loss/crossentropy": 2.2357093393802643,
+      "loss/hidden": 2.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23682421892881395,
+      "loss/reg": 0.0,
+      "step": 13490
+    },
+    {
+      "epoch": 0.08881578947368421,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.05452372233072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1246,
+      "loss/crossentropy": 2.3712179183959963,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2336268275976181,
+      "loss/reg": 0.0,
+      "step": 13500
+    },
+    {
+      "epoch": 0.08888157894736842,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.13904520670572917,
+      "learning_rate": 0.0001,
+      "loss": 3.1879,
+      "loss/crossentropy": 2.407870662212372,
+      "loss/hidden": 2.925,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28650608360767366,
+      "loss/reg": 0.0,
+      "step": 13510
+    },
+    {
+      "epoch": 0.08894736842105264,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.009284464518229167,
+      "learning_rate": 0.0001,
+      "loss": 3.1881,
+      "loss/crossentropy": 2.4734713077545165,
+      "loss/hidden": 3.071875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30961792171001434,
+      "loss/reg": 0.0,
+      "step": 13520
+    },
+    {
+      "epoch": 0.08901315789473684,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.03455301920572917,
+      "learning_rate": 0.0001,
+      "loss": 3.1521,
+      "loss/crossentropy": 2.1405319690704347,
+      "loss/hidden": 2.984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25160129070281984,
+      "loss/reg": 0.0,
+      "step": 13530
+    },
+    {
+      "epoch": 0.08907894736842105,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.13059488932291666,
+      "learning_rate": 0.0001,
+      "loss": 3.2623,
+      "loss/crossentropy": 2.5234264612197874,
+      "loss/hidden": 2.8140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2405148908495903,
+      "loss/reg": 0.0,
+      "step": 13540
+    },
+    {
+      "epoch": 0.08914473684210526,
+      "grad_norm": 3.109375,
+      "grad_norm_var": 0.14059244791666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1637,
+      "loss/crossentropy": 2.4861895561218263,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2647073075175285,
+      "loss/reg": 0.0,
+      "step": 13550
+    },
+    {
+      "epoch": 0.08921052631578948,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.08899332682291666,
+      "learning_rate": 0.0001,
+      "loss": 3.2354,
+      "loss/crossentropy": 2.241242003440857,
+      "loss/hidden": 2.971875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25270578265190125,
+      "loss/reg": 0.0,
+      "step": 13560
+    },
+    {
+      "epoch": 0.08927631578947369,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.15600484212239582,
+      "learning_rate": 0.0001,
+      "loss": 3.2692,
+      "loss/crossentropy": 2.5087321639060973,
+      "loss/hidden": 3.1078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2870207831263542,
+      "loss/reg": 0.0,
+      "step": 13570
+    },
+    {
+      "epoch": 0.08934210526315789,
+      "grad_norm": 1.9140625,
+      "grad_norm_var": 0.09006322224934896,
+      "learning_rate": 0.0001,
+      "loss": 3.2167,
+      "loss/crossentropy": 2.519565200805664,
+      "loss/hidden": 2.9625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2802609995007515,
+      "loss/reg": 0.0,
+      "step": 13580
+    },
+    {
+      "epoch": 0.0894078947368421,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.08097508748372396,
+      "learning_rate": 0.0001,
+      "loss": 3.2167,
+      "loss/crossentropy": 2.6469456434249876,
+      "loss/hidden": 2.83125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2738121926784515,
+      "loss/reg": 0.0,
+      "step": 13590
+    },
+    {
+      "epoch": 0.08947368421052632,
+      "grad_norm": 4.34375,
+      "grad_norm_var": 0.28693033854166666,
+      "learning_rate": 0.0001,
+      "loss": 3.0727,
+      "loss/crossentropy": 2.4280938267707826,
+      "loss/hidden": 2.8875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24971676170825957,
+      "loss/reg": 0.0,
+      "step": 13600
+    },
+    {
+      "epoch": 0.08953947368421053,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.28564046223958334,
+      "learning_rate": 0.0001,
+      "loss": 3.2647,
+      "loss/crossentropy": 2.331229364871979,
+      "loss/hidden": 2.9,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25523588955402376,
+      "loss/reg": 0.0,
+      "step": 13610
+    },
+    {
+      "epoch": 0.08960526315789474,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.1018463134765625,
+      "learning_rate": 0.0001,
+      "loss": 3.141,
+      "loss/crossentropy": 2.248876082897186,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27915321439504626,
+      "loss/reg": 0.0,
+      "step": 13620
+    },
+    {
+      "epoch": 0.08967105263157894,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.05377604166666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1282,
+      "loss/crossentropy": 2.475773072242737,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26457071453332903,
+      "loss/reg": 0.0,
+      "step": 13630
+    },
+    {
+      "epoch": 0.08973684210526316,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.09758707682291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1607,
+      "loss/crossentropy": 2.3480275869369507,
+      "loss/hidden": 2.9828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2902192771434784,
+      "loss/reg": 0.0,
+      "step": 13640
+    },
+    {
+      "epoch": 0.08980263157894737,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.1151519775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1651,
+      "loss/crossentropy": 2.362102711200714,
+      "loss/hidden": 2.96875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24841166138648987,
+      "loss/reg": 0.0,
+      "step": 13650
+    },
+    {
+      "epoch": 0.08986842105263158,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.0971832275390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1725,
+      "loss/crossentropy": 2.134939956665039,
+      "loss/hidden": 2.934375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2594160199165344,
+      "loss/reg": 0.0,
+      "step": 13660
+    },
+    {
+      "epoch": 0.08993421052631578,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.09054361979166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0604,
+      "loss/crossentropy": 2.22348096370697,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2797392845153809,
+      "loss/reg": 0.0,
+      "step": 13670
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.0264556884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.165,
+      "loss/crossentropy": 2.321420121192932,
+      "loss/hidden": 2.95,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25273988842964173,
+      "loss/reg": 0.0,
+      "step": 13680
+    },
+    {
+      "epoch": 0.09006578947368421,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.0448394775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.2251,
+      "loss/crossentropy": 2.5713982224464416,
+      "loss/hidden": 3.1890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3743001103401184,
+      "loss/reg": 0.0,
+      "step": 13690
+    },
+    {
+      "epoch": 0.09013157894736842,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.05821024576822917,
+      "learning_rate": 0.0001,
+      "loss": 3.1174,
+      "loss/crossentropy": 2.433469843864441,
+      "loss/hidden": 2.921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29491184949874877,
+      "loss/reg": 0.0,
+      "step": 13700
+    },
+    {
+      "epoch": 0.09019736842105264,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.06334228515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1422,
+      "loss/crossentropy": 2.498783230781555,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2630519106984138,
+      "loss/reg": 0.0,
+      "step": 13710
+    },
+    {
+      "epoch": 0.09026315789473684,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.18053385416666667,
+      "learning_rate": 0.0001,
+      "loss": 3.2017,
+      "loss/crossentropy": 2.295622777938843,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25550642013549807,
+      "loss/reg": 0.0,
+      "step": 13720
+    },
+    {
+      "epoch": 0.09032894736842105,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.14754231770833334,
+      "learning_rate": 0.0001,
+      "loss": 3.2136,
+      "loss/crossentropy": 2.4960160851478577,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29100794196128843,
+      "loss/reg": 0.0,
+      "step": 13730
+    },
+    {
+      "epoch": 0.09039473684210526,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.060888671875,
+      "learning_rate": 0.0001,
+      "loss": 3.182,
+      "loss/crossentropy": 2.543604516983032,
+      "loss/hidden": 2.9171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26428850889205935,
+      "loss/reg": 0.0,
+      "step": 13740
+    },
+    {
+      "epoch": 0.09046052631578948,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.1883697509765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1144,
+      "loss/crossentropy": 2.495186424255371,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24936733841896058,
+      "loss/reg": 0.0,
+      "step": 13750
+    },
+    {
+      "epoch": 0.09052631578947369,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.17746480305989584,
+      "learning_rate": 0.0001,
+      "loss": 3.2332,
+      "loss/crossentropy": 2.5168472051620485,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3214556619524956,
+      "loss/reg": 0.0,
+      "step": 13760
+    },
+    {
+      "epoch": 0.09059210526315789,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.049853515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1549,
+      "loss/crossentropy": 2.2676196336746215,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28792900443077085,
+      "loss/reg": 0.0,
+      "step": 13770
+    },
+    {
+      "epoch": 0.0906578947368421,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.50240478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1942,
+      "loss/crossentropy": 2.3297463774681093,
+      "loss/hidden": 2.7453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20883744955062866,
+      "loss/reg": 0.0,
+      "step": 13780
+    },
+    {
+      "epoch": 0.09072368421052632,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.4667154947916667,
+      "learning_rate": 0.0001,
+      "loss": 3.2349,
+      "loss/crossentropy": 2.1975256204605103,
+      "loss/hidden": 2.9546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2709361925721169,
+      "loss/reg": 0.0,
+      "step": 13790
+    },
+    {
+      "epoch": 0.09078947368421053,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.03693745930989583,
+      "learning_rate": 0.0001,
+      "loss": 3.1167,
+      "loss/crossentropy": 2.3970743119716644,
+      "loss/hidden": 2.71875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2433250866830349,
+      "loss/reg": 0.0,
+      "step": 13800
+    },
+    {
+      "epoch": 0.09085526315789473,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.7566965738932292,
+      "learning_rate": 0.0001,
+      "loss": 3.2493,
+      "loss/crossentropy": 2.269898569583893,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23498842120170593,
+      "loss/reg": 0.0,
+      "step": 13810
+    },
+    {
+      "epoch": 0.09092105263157894,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.050389607747395836,
+      "learning_rate": 0.0001,
+      "loss": 3.1946,
+      "loss/crossentropy": 2.5624868392944338,
+      "loss/hidden": 3.0046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36430184692144396,
+      "loss/reg": 0.0,
+      "step": 13820
+    },
+    {
+      "epoch": 0.09098684210526316,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.06546122233072917,
+      "learning_rate": 0.0001,
+      "loss": 3.225,
+      "loss/crossentropy": 2.2082170367240908,
+      "loss/hidden": 2.7453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2326130375266075,
+      "loss/reg": 0.0,
+      "step": 13830
+    },
+    {
+      "epoch": 0.09105263157894737,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.1621002197265625,
+      "learning_rate": 0.0001,
+      "loss": 3.225,
+      "loss/crossentropy": 2.2408367514610292,
+      "loss/hidden": 2.9375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25092404931783674,
+      "loss/reg": 0.0,
+      "step": 13840
+    },
+    {
+      "epoch": 0.09111842105263158,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.18394775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1397,
+      "loss/crossentropy": 2.092372101545334,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25034971833229064,
+      "loss/reg": 0.0,
+      "step": 13850
+    },
+    {
+      "epoch": 0.09118421052631578,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.0753313700358073,
+      "learning_rate": 0.0001,
+      "loss": 3.1936,
+      "loss/crossentropy": 2.2277904510498048,
+      "loss/hidden": 3.075,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28555874079465865,
+      "loss/reg": 0.0,
+      "step": 13860
+    },
+    {
+      "epoch": 0.09125,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.06499608357747395,
+      "learning_rate": 0.0001,
+      "loss": 3.1857,
+      "loss/crossentropy": 2.0974882781505584,
+      "loss/hidden": 2.7265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22869385927915573,
+      "loss/reg": 0.0,
+      "step": 13870
+    },
+    {
+      "epoch": 0.09131578947368421,
+      "grad_norm": 3.1875,
+      "grad_norm_var": 0.30895894368489585,
+      "learning_rate": 0.0001,
+      "loss": 3.1808,
+      "loss/crossentropy": 2.2242319107055666,
+      "loss/hidden": 2.925,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24477246254682541,
+      "loss/reg": 0.0,
+      "step": 13880
+    },
+    {
+      "epoch": 0.09138157894736842,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.30500895182291665,
+      "learning_rate": 0.0001,
+      "loss": 3.1587,
+      "loss/crossentropy": 2.2787875294685365,
+      "loss/hidden": 2.8625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23667097985744476,
+      "loss/reg": 0.0,
+      "step": 13890
+    },
+    {
+      "epoch": 0.09144736842105264,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.04114176432291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1791,
+      "loss/crossentropy": 2.5015464782714845,
+      "loss/hidden": 2.9375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26709298938512804,
+      "loss/reg": 0.0,
+      "step": 13900
+    },
+    {
+      "epoch": 0.09151315789473684,
+      "grad_norm": 1.9765625,
+      "grad_norm_var": 0.38769505818684896,
+      "learning_rate": 0.0001,
+      "loss": 3.1858,
+      "loss/crossentropy": 2.367018985748291,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2575811371207237,
+      "loss/reg": 0.0,
+      "step": 13910
+    },
+    {
+      "epoch": 0.09157894736842105,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.12981338500976564,
+      "learning_rate": 0.0001,
+      "loss": 3.2014,
+      "loss/crossentropy": 2.3600114941596986,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2798996135592461,
+      "loss/reg": 0.0,
+      "step": 13920
+    },
+    {
+      "epoch": 0.09164473684210526,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.5571940104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1415,
+      "loss/crossentropy": 2.07312273979187,
+      "loss/hidden": 3.0125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27816066443920134,
+      "loss/reg": 0.0,
+      "step": 13930
+    },
+    {
+      "epoch": 0.09171052631578948,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.16337483723958332,
+      "learning_rate": 0.0001,
+      "loss": 3.1846,
+      "loss/crossentropy": 2.3842572927474976,
+      "loss/hidden": 3.0015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3044495239853859,
+      "loss/reg": 0.0,
+      "step": 13940
+    },
+    {
+      "epoch": 0.09177631578947368,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.011617024739583334,
+      "learning_rate": 0.0001,
+      "loss": 3.2638,
+      "loss/crossentropy": 2.258384811878204,
+      "loss/hidden": 2.9,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24848029464483262,
+      "loss/reg": 0.0,
+      "step": 13950
+    },
+    {
+      "epoch": 0.09184210526315789,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.08046875,
+      "learning_rate": 0.0001,
+      "loss": 3.2277,
+      "loss/crossentropy": 2.1830771923065186,
+      "loss/hidden": 2.95625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.282887265086174,
+      "loss/reg": 0.0,
+      "step": 13960
+    },
+    {
+      "epoch": 0.0919078947368421,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.69127197265625,
+      "learning_rate": 0.0001,
+      "loss": 3.2795,
+      "loss/crossentropy": 2.5693406105041503,
+      "loss/hidden": 3.003125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2618736609816551,
+      "loss/reg": 0.0,
+      "step": 13970
+    },
+    {
+      "epoch": 0.09197368421052632,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 1.2373443603515626,
+      "learning_rate": 0.0001,
+      "loss": 3.163,
+      "loss/crossentropy": 2.496062994003296,
+      "loss/hidden": 2.8625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27756498008966446,
+      "loss/reg": 0.0,
+      "step": 13980
+    },
+    {
+      "epoch": 0.09203947368421053,
+      "grad_norm": 2.375,
+      "grad_norm_var": 1.256787109375,
+      "learning_rate": 0.0001,
+      "loss": 3.2974,
+      "loss/crossentropy": 2.233779698610306,
+      "loss/hidden": 2.8625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22414856255054474,
+      "loss/reg": 0.0,
+      "step": 13990
+    },
+    {
+      "epoch": 0.09210526315789473,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.04104817708333333,
+      "learning_rate": 0.0001,
+      "loss": 3.1465,
+      "loss/crossentropy": 2.3837480187416076,
+      "loss/hidden": 2.8890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29815254360437393,
+      "loss/reg": 0.0,
+      "step": 14000
+    },
+    {
+      "epoch": 0.09217105263157895,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.03504130045572917,
+      "learning_rate": 0.0001,
+      "loss": 3.1933,
+      "loss/crossentropy": 2.4089162349700928,
+      "loss/hidden": 2.940625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28769134283065795,
+      "loss/reg": 0.0,
+      "step": 14010
+    },
+    {
+      "epoch": 0.09223684210526316,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.099560546875,
+      "learning_rate": 0.0001,
+      "loss": 3.2431,
+      "loss/crossentropy": 2.387049177289009,
+      "loss/hidden": 3.078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31867421939969065,
+      "loss/reg": 0.0,
+      "step": 14020
+    },
+    {
+      "epoch": 0.09230263157894737,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.12417704264322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0959,
+      "loss/crossentropy": 2.2009261429309843,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2346379891037941,
+      "loss/reg": 0.0,
+      "step": 14030
+    },
+    {
+      "epoch": 0.09236842105263159,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.0412750244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1679,
+      "loss/crossentropy": 2.379330587387085,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23624493330717086,
+      "loss/reg": 0.0,
+      "step": 14040
+    },
+    {
+      "epoch": 0.09243421052631579,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.25758056640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1901,
+      "loss/crossentropy": 2.2017428398132326,
+      "loss/hidden": 2.9328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2819879144430161,
+      "loss/reg": 0.0,
+      "step": 14050
+    },
+    {
+      "epoch": 0.0925,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 1.2106597900390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1627,
+      "loss/crossentropy": 2.053563690185547,
+      "loss/hidden": 2.8796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24388336688280104,
+      "loss/reg": 0.0,
+      "step": 14060
+    },
+    {
+      "epoch": 0.09256578947368421,
+      "grad_norm": 1.953125,
+      "grad_norm_var": 0.0501617431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1122,
+      "loss/crossentropy": 2.5748242855072023,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2592714115977287,
+      "loss/reg": 0.0,
+      "step": 14070
+    },
+    {
+      "epoch": 0.09263157894736843,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.115869140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2403,
+      "loss/crossentropy": 2.1434300899505616,
+      "loss/hidden": 2.8140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23879239857196807,
+      "loss/reg": 0.0,
+      "step": 14080
+    },
+    {
+      "epoch": 0.09269736842105263,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.08590087890625,
+      "learning_rate": 0.0001,
+      "loss": 3.187,
+      "loss/crossentropy": 2.078695094585419,
+      "loss/hidden": 2.99375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29750160723924635,
+      "loss/reg": 0.0,
+      "step": 14090
+    },
+    {
+      "epoch": 0.09276315789473684,
+      "grad_norm": 3.078125,
+      "grad_norm_var": 0.06355692545572916,
+      "learning_rate": 0.0001,
+      "loss": 3.1426,
+      "loss/crossentropy": 2.4077930808067323,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2784098371863365,
+      "loss/reg": 0.0,
+      "step": 14100
+    },
+    {
+      "epoch": 0.09282894736842105,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.28290913899739584,
+      "learning_rate": 0.0001,
+      "loss": 3.1999,
+      "loss/crossentropy": 2.551260459423065,
+      "loss/hidden": 2.98125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30844325572252274,
+      "loss/reg": 0.0,
+      "step": 14110
+    },
+    {
+      "epoch": 0.09289473684210527,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.28843994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1903,
+      "loss/crossentropy": 2.057539927959442,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21393984705209732,
+      "loss/reg": 0.0,
+      "step": 14120
+    },
+    {
+      "epoch": 0.09296052631578948,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.16035054524739584,
+      "learning_rate": 0.0001,
+      "loss": 3.2036,
+      "loss/crossentropy": 2.2166428923606873,
+      "loss/hidden": 3.009375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28170192539691924,
+      "loss/reg": 0.0,
+      "step": 14130
+    },
+    {
+      "epoch": 0.09302631578947368,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.04431864420572917,
+      "learning_rate": 0.0001,
+      "loss": 3.1706,
+      "loss/crossentropy": 2.5224907636642455,
+      "loss/hidden": 2.8953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2656204476952553,
+      "loss/reg": 0.0,
+      "step": 14140
+    },
+    {
+      "epoch": 0.09309210526315789,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.13594462076822916,
+      "learning_rate": 0.0001,
+      "loss": 3.2365,
+      "loss/crossentropy": 2.365175998210907,
+      "loss/hidden": 2.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28030987083911896,
+      "loss/reg": 0.0,
+      "step": 14150
+    },
+    {
+      "epoch": 0.0931578947368421,
+      "grad_norm": 2.375,
+      "grad_norm_var": 3.4480684566712595e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.3698,
+      "loss/crossentropy": 2.4446977496147158,
+      "loss/hidden": 4.159375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35163595527410507,
+      "loss/reg": 0.0,
+      "step": 14160
+    },
+    {
+      "epoch": 0.09322368421052632,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 3.4480684570076774e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1773,
+      "loss/crossentropy": 2.144097054004669,
+      "loss/hidden": 3.0171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2777150124311447,
+      "loss/reg": 0.0,
+      "step": 14170
+    },
+    {
+      "epoch": 0.09328947368421053,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.1343902587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1593,
+      "loss/crossentropy": 2.5162782430648805,
+      "loss/hidden": 2.884375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2642303377389908,
+      "loss/reg": 0.0,
+      "step": 14180
+    },
+    {
+      "epoch": 0.09335526315789473,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.169921875,
+      "learning_rate": 0.0001,
+      "loss": 3.1927,
+      "loss/crossentropy": 2.1481271982192993,
+      "loss/hidden": 3.034375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27440374791622163,
+      "loss/reg": 0.0,
+      "step": 14190
+    },
+    {
+      "epoch": 0.09342105263157895,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.188330078125,
+      "learning_rate": 0.0001,
+      "loss": 3.2049,
+      "loss/crossentropy": 2.4672377467155457,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25575320422649384,
+      "loss/reg": 0.0,
+      "step": 14200
+    },
+    {
+      "epoch": 0.09348684210526316,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.07030843098958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1256,
+      "loss/crossentropy": 2.4822991728782653,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23255562111735345,
+      "loss/reg": 0.0,
+      "step": 14210
+    },
+    {
+      "epoch": 0.09355263157894737,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.06575698852539062,
+      "learning_rate": 0.0001,
+      "loss": 3.1061,
+      "loss/crossentropy": 2.4455429315567017,
+      "loss/hidden": 2.7171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2238618478178978,
+      "loss/reg": 0.0,
+      "step": 14220
+    },
+    {
+      "epoch": 0.09361842105263157,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.13621317545572917,
+      "learning_rate": 0.0001,
+      "loss": 3.1927,
+      "loss/crossentropy": 2.3609827399253844,
+      "loss/hidden": 3.021875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3211557373404503,
+      "loss/reg": 0.0,
+      "step": 14230
+    },
+    {
+      "epoch": 0.09368421052631579,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.1390777587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.2222,
+      "loss/crossentropy": 2.4929265141487122,
+      "loss/hidden": 2.921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.256229493021965,
+      "loss/reg": 0.0,
+      "step": 14240
+    },
+    {
+      "epoch": 0.09375,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.014774576822916666,
+      "learning_rate": 0.0001,
+      "loss": 3.0973,
+      "loss/crossentropy": 2.396563506126404,
+      "loss/hidden": 2.9515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3292677074670792,
+      "loss/reg": 0.0,
+      "step": 14250
+    },
+    {
+      "epoch": 0.09381578947368421,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 3.801495474913411e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.3636,
+      "loss/crossentropy": 2.1659668326377868,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24123398140072821,
+      "loss/reg": 0.0,
+      "step": 14260
+    },
+    {
+      "epoch": 0.09388157894736843,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 3.801495474059215e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2259,
+      "loss/crossentropy": 2.35439647436142,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22794780433177947,
+      "loss/reg": 0.0,
+      "step": 14270
+    },
+    {
+      "epoch": 0.09394736842105263,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.42942708333333335,
+      "learning_rate": 0.0001,
+      "loss": 3.1665,
+      "loss/crossentropy": 2.3001658797264097,
+      "loss/hidden": 2.984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26924325078725814,
+      "loss/reg": 0.0,
+      "step": 14280
+    },
+    {
+      "epoch": 0.09401315789473684,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.0408843994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0518,
+      "loss/crossentropy": 2.3841129422187803,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26079314202070236,
+      "loss/reg": 0.0,
+      "step": 14290
+    },
+    {
+      "epoch": 0.09407894736842105,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.03784891764322917,
+      "learning_rate": 0.0001,
+      "loss": 3.1904,
+      "loss/crossentropy": 2.2710848689079284,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23626017868518828,
+      "loss/reg": 0.0,
+      "step": 14300
+    },
+    {
+      "epoch": 0.09414473684210527,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.11845296223958333,
+      "learning_rate": 0.0001,
+      "loss": 3.2715,
+      "loss/crossentropy": 2.4462394237518312,
+      "loss/hidden": 3.0515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3083674684166908,
+      "loss/reg": 0.0,
+      "step": 14310
+    },
+    {
+      "epoch": 0.09421052631578947,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.115771484375,
+      "learning_rate": 0.0001,
+      "loss": 3.1137,
+      "loss/crossentropy": 2.4102493643760683,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3285558119416237,
+      "loss/reg": 0.0,
+      "step": 14320
+    },
+    {
+      "epoch": 0.09427631578947368,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.0410552978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1163,
+      "loss/crossentropy": 2.496846008300781,
+      "loss/hidden": 2.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2681970477104187,
+      "loss/reg": 0.0,
+      "step": 14330
+    },
+    {
+      "epoch": 0.0943421052631579,
+      "grad_norm": 3.28125,
+      "grad_norm_var": 0.3074615478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.2179,
+      "loss/crossentropy": 2.368880546092987,
+      "loss/hidden": 2.9671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27040198296308515,
+      "loss/reg": 0.0,
+      "step": 14340
+    },
+    {
+      "epoch": 0.09440789473684211,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.087841796875,
+      "learning_rate": 0.0001,
+      "loss": 3.1677,
+      "loss/crossentropy": 2.379721689224243,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3262082427740097,
+      "loss/reg": 0.0,
+      "step": 14350
+    },
+    {
+      "epoch": 0.09447368421052632,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.06515299479166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1541,
+      "loss/crossentropy": 2.3577764987945558,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28836628049612045,
+      "loss/reg": 0.0,
+      "step": 14360
+    },
+    {
+      "epoch": 0.09453947368421052,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.09937744140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1429,
+      "loss/crossentropy": 2.2929248332977297,
+      "loss/hidden": 2.9015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2473236471414566,
+      "loss/reg": 0.0,
+      "step": 14370
+    },
+    {
+      "epoch": 0.09460526315789473,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.056428019205729166,
+      "learning_rate": 0.0001,
+      "loss": 3.1331,
+      "loss/crossentropy": 2.11824688911438,
+      "loss/hidden": 2.59375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19864632040262223,
+      "loss/reg": 0.0,
+      "step": 14380
+    },
+    {
+      "epoch": 0.09467105263157895,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.2792154947916667,
+      "learning_rate": 0.0001,
+      "loss": 3.2558,
+      "loss/crossentropy": 2.4552414536476137,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26964131295681,
+      "loss/reg": 0.0,
+      "step": 14390
+    },
+    {
+      "epoch": 0.09473684210526316,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.1197418212890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1627,
+      "loss/crossentropy": 2.4108232736587523,
+      "loss/hidden": 2.93125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2856267586350441,
+      "loss/reg": 0.0,
+      "step": 14400
+    },
+    {
+      "epoch": 0.09480263157894737,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.09917704264322917,
+      "learning_rate": 0.0001,
+      "loss": 3.2683,
+      "loss/crossentropy": 2.245331883430481,
+      "loss/hidden": 2.9640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26227104514837263,
+      "loss/reg": 0.0,
+      "step": 14410
+    },
+    {
+      "epoch": 0.09486842105263157,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.47395426432291665,
+      "learning_rate": 0.0001,
+      "loss": 3.3242,
+      "loss/crossentropy": 2.5019222021102907,
+      "loss/hidden": 2.8578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2487493023276329,
+      "loss/reg": 0.0,
+      "step": 14420
+    },
+    {
+      "epoch": 0.09493421052631579,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.8088053385416667,
+      "learning_rate": 0.0001,
+      "loss": 3.1846,
+      "loss/crossentropy": 2.239874541759491,
+      "loss/hidden": 2.9359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30340075492858887,
+      "loss/reg": 0.0,
+      "step": 14430
+    },
+    {
+      "epoch": 0.095,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.5774698893229167,
+      "learning_rate": 0.0001,
+      "loss": 3.2356,
+      "loss/crossentropy": 2.212349569797516,
+      "loss/hidden": 2.9453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25182478949427606,
+      "loss/reg": 0.0,
+      "step": 14440
+    },
+    {
+      "epoch": 0.09506578947368421,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.19168294270833333,
+      "learning_rate": 0.0001,
+      "loss": 3.2659,
+      "loss/crossentropy": 2.337146294116974,
+      "loss/hidden": 2.9328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2963370710611343,
+      "loss/reg": 0.0,
+      "step": 14450
+    },
+    {
+      "epoch": 0.09513157894736841,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.10920308430989584,
+      "learning_rate": 0.0001,
+      "loss": 3.2442,
+      "loss/crossentropy": 1.8299875736236573,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21587850898504257,
+      "loss/reg": 0.0,
+      "step": 14460
+    },
+    {
+      "epoch": 0.09519736842105263,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.04309794108072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1964,
+      "loss/crossentropy": 2.558639335632324,
+      "loss/hidden": 2.828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25251512974500656,
+      "loss/reg": 0.0,
+      "step": 14470
+    },
+    {
+      "epoch": 0.09526315789473684,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.020686848958333334,
+      "learning_rate": 0.0001,
+      "loss": 3.1942,
+      "loss/crossentropy": 2.514283466339111,
+      "loss/hidden": 3.1234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2890095472335815,
+      "loss/reg": 0.0,
+      "step": 14480
+    },
+    {
+      "epoch": 0.09532894736842105,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.028123982747395835,
+      "learning_rate": 0.0001,
+      "loss": 3.1856,
+      "loss/crossentropy": 2.2959898948669433,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24410185664892198,
+      "loss/reg": 0.0,
+      "step": 14490
+    },
+    {
+      "epoch": 0.09539473684210527,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.03728841145833333,
+      "learning_rate": 0.0001,
+      "loss": 3.2528,
+      "loss/crossentropy": 2.421563959121704,
+      "loss/hidden": 3.0015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30335861891508104,
+      "loss/reg": 0.0,
+      "step": 14500
+    },
+    {
+      "epoch": 0.09546052631578947,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.041825358072916666,
+      "learning_rate": 0.0001,
+      "loss": 3.1292,
+      "loss/crossentropy": 2.3713988065719604,
+      "loss/hidden": 2.884375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2508593872189522,
+      "loss/reg": 0.0,
+      "step": 14510
+    },
+    {
+      "epoch": 0.09552631578947368,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.12280985514322916,
+      "learning_rate": 0.0001,
+      "loss": 3.1975,
+      "loss/crossentropy": 2.3875895380973815,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26996816471219065,
+      "loss/reg": 0.0,
+      "step": 14520
+    },
+    {
+      "epoch": 0.0955921052631579,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.13038101196289062,
+      "learning_rate": 0.0001,
+      "loss": 3.1709,
+      "loss/crossentropy": 2.3352912187576296,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23363660275936127,
+      "loss/reg": 0.0,
+      "step": 14530
+    },
+    {
+      "epoch": 0.09565789473684211,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.11688206990559896,
+      "learning_rate": 0.0001,
+      "loss": 3.1364,
+      "loss/crossentropy": 2.473167669773102,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24549597799777984,
+      "loss/reg": 0.0,
+      "step": 14540
+    },
+    {
+      "epoch": 0.09572368421052632,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.060868326822916666,
+      "learning_rate": 0.0001,
+      "loss": 3.2086,
+      "loss/crossentropy": 2.3421871423721314,
+      "loss/hidden": 2.890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2576961562037468,
+      "loss/reg": 0.0,
+      "step": 14550
+    },
+    {
+      "epoch": 0.09578947368421052,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.10832926432291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1744,
+      "loss/crossentropy": 2.1765161633491514,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2244855582714081,
+      "loss/reg": 0.0,
+      "step": 14560
+    },
+    {
+      "epoch": 0.09585526315789474,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.1071685791015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1489,
+      "loss/crossentropy": 2.507940888404846,
+      "loss/hidden": 2.9671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2875757083296776,
+      "loss/reg": 0.0,
+      "step": 14570
+    },
+    {
+      "epoch": 0.09592105263157895,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.029523722330729165,
+      "learning_rate": 0.0001,
+      "loss": 3.1669,
+      "loss/crossentropy": 2.2886616230010985,
+      "loss/hidden": 3.0765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3050649344921112,
+      "loss/reg": 0.0,
+      "step": 14580
+    },
+    {
+      "epoch": 0.09598684210526316,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.03560282389322917,
+      "learning_rate": 0.0001,
+      "loss": 3.1454,
+      "loss/crossentropy": 2.395359969139099,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2635947346687317,
+      "loss/reg": 0.0,
+      "step": 14590
+    },
+    {
+      "epoch": 0.09605263157894736,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.23357747395833334,
+      "learning_rate": 0.0001,
+      "loss": 3.2176,
+      "loss/crossentropy": 2.002879011631012,
+      "loss/hidden": 2.9875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24021245390176774,
+      "loss/reg": 0.0,
+      "step": 14600
+    },
+    {
+      "epoch": 0.09611842105263158,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.216552734375,
+      "learning_rate": 0.0001,
+      "loss": 3.1289,
+      "loss/crossentropy": 2.4440099954605103,
+      "loss/hidden": 3.0421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3085349440574646,
+      "loss/reg": 0.0,
+      "step": 14610
+    },
+    {
+      "epoch": 0.09618421052631579,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.18892822265625,
+      "learning_rate": 0.0001,
+      "loss": 3.2541,
+      "loss/crossentropy": 2.5584524154663084,
+      "loss/hidden": 2.978125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33079054951667786,
+      "loss/reg": 0.0,
+      "step": 14620
+    },
+    {
+      "epoch": 0.09625,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.18036702473958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1899,
+      "loss/crossentropy": 2.386956262588501,
+      "loss/hidden": 2.9546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2895423695445061,
+      "loss/reg": 0.0,
+      "step": 14630
+    },
+    {
+      "epoch": 0.09631578947368422,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.15852762858072916,
+      "learning_rate": 0.0001,
+      "loss": 3.1443,
+      "loss/crossentropy": 2.0192247331142426,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2274337187409401,
+      "loss/reg": 0.0,
+      "step": 14640
+    },
+    {
+      "epoch": 0.09638157894736842,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.11549072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.1243,
+      "loss/crossentropy": 2.499123454093933,
+      "loss/hidden": 2.828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25674946457147596,
+      "loss/reg": 0.0,
+      "step": 14650
+    },
+    {
+      "epoch": 0.09644736842105263,
+      "grad_norm": 2130706432.0,
+      "grad_norm_var": 2.83744368059244e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.251,
+      "loss/crossentropy": 2.1218835711479187,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23685040771961213,
+      "loss/reg": 0.0,
+      "step": 14660
+    },
+    {
+      "epoch": 0.09651315789473684,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 2.8374436794771485e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2008,
+      "loss/crossentropy": 2.125733083486557,
+      "loss/hidden": 3.0,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3448520749807358,
+      "loss/reg": 0.0,
+      "step": 14670
+    },
+    {
+      "epoch": 0.09657894736842106,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.24964090983072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1553,
+      "loss/crossentropy": 2.604245328903198,
+      "loss/hidden": 2.871875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23915667235851287,
+      "loss/reg": 0.0,
+      "step": 14680
+    },
+    {
+      "epoch": 0.09664473684210527,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.11900634765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1023,
+      "loss/crossentropy": 2.3761191368103027,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2546194761991501,
+      "loss/reg": 0.0,
+      "step": 14690
+    },
+    {
+      "epoch": 0.09671052631578947,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.05559488932291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1544,
+      "loss/crossentropy": 2.3285223722457884,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24710593670606612,
+      "loss/reg": 0.0,
+      "step": 14700
+    },
+    {
+      "epoch": 0.09677631578947368,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.13647842407226562,
+      "learning_rate": 0.0001,
+      "loss": 3.1385,
+      "loss/crossentropy": 2.364056706428528,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22887694388628005,
+      "loss/reg": 0.0,
+      "step": 14710
+    },
+    {
+      "epoch": 0.0968421052631579,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 5.376778157552083,
+      "learning_rate": 0.0001,
+      "loss": 3.1773,
+      "loss/crossentropy": 2.2852516174316406,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24893905222415924,
+      "loss/reg": 0.0,
+      "step": 14720
+    },
+    {
+      "epoch": 0.09690789473684211,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 5.43084487915039,
+      "learning_rate": 0.0001,
+      "loss": 3.1089,
+      "loss/crossentropy": 2.1680606245994567,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23210474848747253,
+      "loss/reg": 0.0,
+      "step": 14730
+    },
+    {
+      "epoch": 0.09697368421052631,
+      "grad_norm": 1.984375,
+      "grad_norm_var": 0.1650054931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1202,
+      "loss/crossentropy": 2.1676085114479067,
+      "loss/hidden": 3.103125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2939779102802277,
+      "loss/reg": 0.0,
+      "step": 14740
+    },
+    {
+      "epoch": 0.09703947368421052,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.17122294108072916,
+      "learning_rate": 0.0001,
+      "loss": 3.0635,
+      "loss/crossentropy": 2.498277449607849,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22846879661083222,
+      "loss/reg": 0.0,
+      "step": 14750
+    },
+    {
+      "epoch": 0.09710526315789474,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.11088765462239583,
+      "learning_rate": 0.0001,
+      "loss": 3.155,
+      "loss/crossentropy": 2.2534152626991273,
+      "loss/hidden": 2.9390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2535191968083382,
+      "loss/reg": 0.0,
+      "step": 14760
+    },
+    {
+      "epoch": 0.09717105263157895,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.10987955729166667,
+      "learning_rate": 0.0001,
+      "loss": 3.2477,
+      "loss/crossentropy": 2.085835373401642,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25164034962654114,
+      "loss/reg": 0.0,
+      "step": 14770
+    },
+    {
+      "epoch": 0.09723684210526316,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 17.859382120768228,
+      "learning_rate": 0.0001,
+      "loss": 3.2842,
+      "loss/crossentropy": 1.8799749910831451,
+      "loss/hidden": 3.0765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24884901493787764,
+      "loss/reg": 0.0,
+      "step": 14780
+    },
+    {
+      "epoch": 0.09730263157894736,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 17.996207682291665,
+      "learning_rate": 0.0001,
+      "loss": 3.2386,
+      "loss/crossentropy": 2.307372045516968,
+      "loss/hidden": 3.0421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26386404484510423,
+      "loss/reg": 0.0,
+      "step": 14790
+    },
+    {
+      "epoch": 0.09736842105263158,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.06170247395833333,
+      "learning_rate": 0.0001,
+      "loss": 3.1452,
+      "loss/crossentropy": 2.454807901382446,
+      "loss/hidden": 2.9390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3172868087887764,
+      "loss/reg": 0.0,
+      "step": 14800
+    },
+    {
+      "epoch": 0.09743421052631579,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.05920817057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1114,
+      "loss/crossentropy": 2.014724650979042,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1908031925559044,
+      "loss/reg": 0.0,
+      "step": 14810
+    },
+    {
+      "epoch": 0.0975,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.060445149739583336,
+      "learning_rate": 0.0001,
+      "loss": 3.2493,
+      "loss/crossentropy": 2.463605833053589,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24406316578388215,
+      "loss/reg": 0.0,
+      "step": 14820
+    },
+    {
+      "epoch": 0.09756578947368422,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.29640299479166665,
+      "learning_rate": 0.0001,
+      "loss": 3.1614,
+      "loss/crossentropy": 2.501069176197052,
+      "loss/hidden": 2.8453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2565233051776886,
+      "loss/reg": 0.0,
+      "step": 14830
+    },
+    {
+      "epoch": 0.09763157894736842,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.5406534830729167,
+      "learning_rate": 0.0001,
+      "loss": 3.1904,
+      "loss/crossentropy": 2.462419664859772,
+      "loss/hidden": 2.7578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22374519556760789,
+      "loss/reg": 0.0,
+      "step": 14840
+    },
+    {
+      "epoch": 0.09769736842105263,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.7624501546223958,
+      "learning_rate": 0.0001,
+      "loss": 3.2431,
+      "loss/crossentropy": 2.226536822319031,
+      "loss/hidden": 3.1921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26644248366355894,
+      "loss/reg": 0.0,
+      "step": 14850
+    },
+    {
+      "epoch": 0.09776315789473684,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 2.5783355712890623,
+      "learning_rate": 0.0001,
+      "loss": 3.3229,
+      "loss/crossentropy": 2.3811925053596497,
+      "loss/hidden": 2.715625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22725498378276826,
+      "loss/reg": 0.0,
+      "step": 14860
+    },
+    {
+      "epoch": 0.09782894736842106,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 2.3144205729166667,
+      "learning_rate": 0.0001,
+      "loss": 3.2061,
+      "loss/crossentropy": 2.5685184717178347,
+      "loss/hidden": 2.828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2757527410984039,
+      "loss/reg": 0.0,
+      "step": 14870
+    },
+    {
+      "epoch": 0.09789473684210526,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 1.900218709309896,
+      "learning_rate": 0.0001,
+      "loss": 3.1588,
+      "loss/crossentropy": 2.224149799346924,
+      "loss/hidden": 2.95,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24766245782375335,
+      "loss/reg": 0.0,
+      "step": 14880
+    },
+    {
+      "epoch": 0.09796052631578947,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.13209228515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1723,
+      "loss/crossentropy": 2.2869945645332335,
+      "loss/hidden": 3.0,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3204286351799965,
+      "loss/reg": 0.0,
+      "step": 14890
+    },
+    {
+      "epoch": 0.09802631578947368,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.036408487955729166,
+      "learning_rate": 0.0001,
+      "loss": 3.1244,
+      "loss/crossentropy": 2.276153302192688,
+      "loss/hidden": 2.753125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2348289594054222,
+      "loss/reg": 0.0,
+      "step": 14900
+    },
+    {
+      "epoch": 0.0980921052631579,
+      "grad_norm": 3.328125,
+      "grad_norm_var": 0.0842926025390625,
+      "learning_rate": 0.0001,
+      "loss": 3.2024,
+      "loss/crossentropy": 2.2277270436286924,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2504657179117203,
+      "loss/reg": 0.0,
+      "step": 14910
+    },
+    {
+      "epoch": 0.09815789473684211,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.15784098307291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1643,
+      "loss/crossentropy": 2.601578450202942,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27564091980457306,
+      "loss/reg": 0.0,
+      "step": 14920
+    },
+    {
+      "epoch": 0.09822368421052631,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.12454325358072917,
+      "learning_rate": 0.0001,
+      "loss": 3.0857,
+      "loss/crossentropy": 2.4973155736923216,
+      "loss/hidden": 2.9234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2681492820382118,
+      "loss/reg": 0.0,
+      "step": 14930
+    },
+    {
+      "epoch": 0.09828947368421052,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.04514872233072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1214,
+      "loss/crossentropy": 2.526623797416687,
+      "loss/hidden": 2.8140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26705596745014193,
+      "loss/reg": 0.0,
+      "step": 14940
+    },
+    {
+      "epoch": 0.09835526315789474,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.23338216145833332,
+      "learning_rate": 0.0001,
+      "loss": 3.1468,
+      "loss/crossentropy": 2.273276376724243,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28430237621068954,
+      "loss/reg": 0.0,
+      "step": 14950
+    },
+    {
+      "epoch": 0.09842105263157895,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.2802398681640625,
+      "learning_rate": 0.0001,
+      "loss": 3.2042,
+      "loss/crossentropy": 2.4497693538665772,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2505017280578613,
+      "loss/reg": 0.0,
+      "step": 14960
+    },
+    {
+      "epoch": 0.09848684210526316,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.15282796223958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1746,
+      "loss/crossentropy": 2.4736087679862977,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24588750153779984,
+      "loss/reg": 0.0,
+      "step": 14970
+    },
+    {
+      "epoch": 0.09855263157894736,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.11015218098958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1539,
+      "loss/crossentropy": 2.373614990711212,
+      "loss/hidden": 3.0578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30631934851408005,
+      "loss/reg": 0.0,
+      "step": 14980
+    },
+    {
+      "epoch": 0.09861842105263158,
+      "grad_norm": 3.921875,
+      "grad_norm_var": 0.1558990478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1982,
+      "loss/crossentropy": 2.295030379295349,
+      "loss/hidden": 2.8203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24311772882938384,
+      "loss/reg": 0.0,
+      "step": 14990
+    },
+    {
+      "epoch": 0.09868421052631579,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.3582967122395833,
+      "learning_rate": 0.0001,
+      "loss": 3.2225,
+      "loss/crossentropy": 2.29546434879303,
+      "loss/hidden": 3.05625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.45574710667133334,
+      "loss/reg": 0.0,
+      "step": 15000
+    },
+    {
+      "epoch": 0.09875,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 5.883512115478515,
+      "learning_rate": 0.0001,
+      "loss": 3.1782,
+      "loss/crossentropy": 2.4186841249465942,
+      "loss/hidden": 3.0265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3855607584118843,
+      "loss/reg": 0.0,
+      "step": 15010
+    },
+    {
+      "epoch": 0.0988157894736842,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 5.986443837483724,
+      "learning_rate": 0.0001,
+      "loss": 3.0661,
+      "loss/crossentropy": 2.409875476360321,
+      "loss/hidden": 2.884375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23775186911225318,
+      "loss/reg": 0.0,
+      "step": 15020
+    },
+    {
+      "epoch": 0.09888157894736842,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.055916086832682295,
+      "learning_rate": 0.0001,
+      "loss": 3.1756,
+      "loss/crossentropy": 2.187470281124115,
+      "loss/hidden": 3.00625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2504939392209053,
+      "loss/reg": 0.0,
+      "step": 15030
+    },
+    {
+      "epoch": 0.09894736842105263,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.058203125,
+      "learning_rate": 0.0001,
+      "loss": 3.2179,
+      "loss/crossentropy": 2.634449529647827,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2524956986308098,
+      "loss/reg": 0.0,
+      "step": 15040
+    },
+    {
+      "epoch": 0.09901315789473684,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.09206441243489584,
+      "learning_rate": 0.0001,
+      "loss": 3.1804,
+      "loss/crossentropy": 2.3493194222450255,
+      "loss/hidden": 2.921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24196315556764603,
+      "loss/reg": 0.0,
+      "step": 15050
+    },
+    {
+      "epoch": 0.09907894736842106,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.0587799072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.1473,
+      "loss/crossentropy": 2.5005852222442626,
+      "loss/hidden": 3.015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2778876781463623,
+      "loss/reg": 0.0,
+      "step": 15060
+    },
+    {
+      "epoch": 0.09914473684210526,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.07665913899739583,
+      "learning_rate": 0.0001,
+      "loss": 3.2621,
+      "loss/crossentropy": 2.30003308057785,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28490780740976335,
+      "loss/reg": 0.0,
+      "step": 15070
+    },
+    {
+      "epoch": 0.09921052631578947,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.07154541015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1472,
+      "loss/crossentropy": 2.2739776968955994,
+      "loss/hidden": 2.975,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2682348355650902,
+      "loss/reg": 0.0,
+      "step": 15080
+    },
+    {
+      "epoch": 0.09927631578947368,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.044169108072916664,
+      "learning_rate": 0.0001,
+      "loss": 3.1016,
+      "loss/crossentropy": 2.2083258867263793,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23891242742538452,
+      "loss/reg": 0.0,
+      "step": 15090
+    },
+    {
+      "epoch": 0.0993421052631579,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.0837890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1153,
+      "loss/crossentropy": 2.4516371846199037,
+      "loss/hidden": 2.821875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21450784876942636,
+      "loss/reg": 0.0,
+      "step": 15100
+    },
+    {
+      "epoch": 0.09940789473684211,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.14197489420572917,
+      "learning_rate": 0.0001,
+      "loss": 3.1489,
+      "loss/crossentropy": 2.2019423693418503,
+      "loss/hidden": 2.925,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.278714843839407,
+      "loss/reg": 0.0,
+      "step": 15110
+    },
+    {
+      "epoch": 0.09947368421052631,
+      "grad_norm": 1.984375,
+      "grad_norm_var": 0.06102676391601562,
+      "learning_rate": 0.0001,
+      "loss": 3.0573,
+      "loss/crossentropy": 2.0795727133750916,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24602137356996537,
+      "loss/reg": 0.0,
+      "step": 15120
+    },
+    {
+      "epoch": 0.09953947368421052,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.04859619140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1684,
+      "loss/crossentropy": 2.0709302008152006,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27385044246912005,
+      "loss/reg": 0.0,
+      "step": 15130
+    },
+    {
+      "epoch": 0.09960526315789474,
+      "grad_norm": 3.359375,
+      "grad_norm_var": 0.10572509765625,
+      "learning_rate": 0.0001,
+      "loss": 3.2319,
+      "loss/crossentropy": 2.1219942808151244,
+      "loss/hidden": 3.0421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31611712723970414,
+      "loss/reg": 0.0,
+      "step": 15140
+    },
+    {
+      "epoch": 0.09967105263157895,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.30730794270833334,
+      "learning_rate": 0.0001,
+      "loss": 3.3146,
+      "loss/crossentropy": 2.36390962600708,
+      "loss/hidden": 2.7359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2202586129307747,
+      "loss/reg": 0.0,
+      "step": 15150
+    },
+    {
+      "epoch": 0.09973684210526315,
+      "grad_norm": 3.015625,
+      "grad_norm_var": 0.3680948893229167,
+      "learning_rate": 0.0001,
+      "loss": 3.1903,
+      "loss/crossentropy": 2.592810094356537,
+      "loss/hidden": 2.959375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31535505652427676,
+      "loss/reg": 0.0,
+      "step": 15160
+    },
+    {
+      "epoch": 0.09980263157894737,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.06611226399739584,
+      "learning_rate": 0.0001,
+      "loss": 3.1583,
+      "loss/crossentropy": 2.4805197715759277,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25672143548727033,
+      "loss/reg": 0.0,
+      "step": 15170
+    },
+    {
+      "epoch": 0.09986842105263158,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.026691691080729166,
+      "learning_rate": 0.0001,
+      "loss": 3.0654,
+      "loss/crossentropy": 2.428369462490082,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2483712613582611,
+      "loss/reg": 0.0,
+      "step": 15180
+    },
+    {
+      "epoch": 0.09993421052631579,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.16633707682291668,
+      "learning_rate": 0.0001,
+      "loss": 3.2011,
+      "loss/crossentropy": 1.8888699412345886,
+      "loss/hidden": 3.059375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25502131283283236,
+      "loss/reg": 0.0,
+      "step": 15190
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.09675191243489584,
+      "learning_rate": 0.0001,
+      "loss": 3.1893,
+      "loss/crossentropy": 2.1418832421302794,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2491638869047165,
+      "loss/reg": 0.0,
+      "step": 15200
+    },
+    {
+      "epoch": 0.1000657894736842,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.08440755208333334,
+      "learning_rate": 0.0001,
+      "loss": 3.2207,
+      "loss/crossentropy": 2.5305609703063965,
+      "loss/hidden": 2.9046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28034003674983976,
+      "loss/reg": 0.0,
+      "step": 15210
+    },
+    {
+      "epoch": 0.10013157894736842,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.10357666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.2279,
+      "loss/crossentropy": 2.1161023378372192,
+      "loss/hidden": 2.975,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2513925403356552,
+      "loss/reg": 0.0,
+      "step": 15220
+    },
+    {
+      "epoch": 0.10019736842105263,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.2027008056640625,
+      "learning_rate": 0.0001,
+      "loss": 3.2361,
+      "loss/crossentropy": 2.064685332775116,
+      "loss/hidden": 2.9890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28131103515625,
+      "loss/reg": 0.0,
+      "step": 15230
+    },
+    {
+      "epoch": 0.10026315789473685,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.14849853515625,
+      "learning_rate": 0.0001,
+      "loss": 3.151,
+      "loss/crossentropy": 2.4768277525901796,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24746784269809724,
+      "loss/reg": 0.0,
+      "step": 15240
+    },
+    {
+      "epoch": 0.10032894736842106,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.07953999837239584,
+      "learning_rate": 0.0001,
+      "loss": 3.1794,
+      "loss/crossentropy": 2.2046443462371825,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26957911550998687,
+      "loss/reg": 0.0,
+      "step": 15250
+    },
+    {
+      "epoch": 0.10039473684210526,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.12280985514322916,
+      "learning_rate": 0.0001,
+      "loss": 3.0606,
+      "loss/crossentropy": 2.418226730823517,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21895384043455124,
+      "loss/reg": 0.0,
+      "step": 15260
+    },
+    {
+      "epoch": 0.10046052631578947,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.09387613932291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1144,
+      "loss/crossentropy": 2.1524213790893554,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23847964853048326,
+      "loss/reg": 0.0,
+      "step": 15270
+    },
+    {
+      "epoch": 0.10052631578947369,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.12705459594726562,
+      "learning_rate": 0.0001,
+      "loss": 3.0626,
+      "loss/crossentropy": 1.9326023817062379,
+      "loss/hidden": 2.946875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24662483483552933,
+      "loss/reg": 0.0,
+      "step": 15280
+    },
+    {
+      "epoch": 0.1005921052631579,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 8.221414947509766,
+      "learning_rate": 0.0001,
+      "loss": 3.2248,
+      "loss/crossentropy": 2.2926568508148195,
+      "loss/hidden": 2.9828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26595802754163744,
+      "loss/reg": 0.0,
+      "step": 15290
+    },
+    {
+      "epoch": 0.1006578947368421,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 8.17893778483073,
+      "learning_rate": 0.0001,
+      "loss": 3.1429,
+      "loss/crossentropy": 2.4002704977989198,
+      "loss/hidden": 3.0109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2911446109414101,
+      "loss/reg": 0.0,
+      "step": 15300
+    },
+    {
+      "epoch": 0.10072368421052631,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.03648681640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1424,
+      "loss/crossentropy": 2.1603388369083403,
+      "loss/hidden": 2.9984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27493633329868317,
+      "loss/reg": 0.0,
+      "step": 15310
+    },
+    {
+      "epoch": 0.10078947368421053,
+      "grad_norm": 3.96875,
+      "grad_norm_var": 0.1912506103515625,
+      "learning_rate": 0.0001,
+      "loss": 3.133,
+      "loss/crossentropy": 2.279471695423126,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23778771311044694,
+      "loss/reg": 0.0,
+      "step": 15320
+    },
+    {
+      "epoch": 0.10085526315789474,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.2115875244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0811,
+      "loss/crossentropy": 2.1736138820648194,
+      "loss/hidden": 2.665625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19810649901628494,
+      "loss/reg": 0.0,
+      "step": 15330
+    },
+    {
+      "epoch": 0.10092105263157895,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.05100809733072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1508,
+      "loss/crossentropy": 2.2725695729255677,
+      "loss/hidden": 3.0609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2860790088772774,
+      "loss/reg": 0.0,
+      "step": 15340
+    },
+    {
+      "epoch": 0.10098684210526315,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 2.056571451822917,
+      "learning_rate": 0.0001,
+      "loss": 3.2468,
+      "loss/crossentropy": 2.215101981163025,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27275211960077284,
+      "loss/reg": 0.0,
+      "step": 15350
+    },
+    {
+      "epoch": 0.10105263157894737,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 2.1172159830729167,
+      "learning_rate": 0.0001,
+      "loss": 3.1519,
+      "loss/crossentropy": 2.416505420207977,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26659835278987887,
+      "loss/reg": 0.0,
+      "step": 15360
+    },
+    {
+      "epoch": 0.10111842105263158,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.051708984375,
+      "learning_rate": 0.0001,
+      "loss": 3.0702,
+      "loss/crossentropy": 2.3915260195732118,
+      "loss/hidden": 2.7171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23197022825479507,
+      "loss/reg": 0.0,
+      "step": 15370
+    },
+    {
+      "epoch": 0.1011842105263158,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.03733622233072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1794,
+      "loss/crossentropy": 2.4574419140815733,
+      "loss/hidden": 3.134375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34623306542634963,
+      "loss/reg": 0.0,
+      "step": 15380
+    },
+    {
+      "epoch": 0.10125,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.229052734375,
+      "learning_rate": 0.0001,
+      "loss": 3.1256,
+      "loss/crossentropy": 2.1932177782058715,
+      "loss/hidden": 2.9078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22062241584062575,
+      "loss/reg": 0.0,
+      "step": 15390
+    },
+    {
+      "epoch": 0.1013157894736842,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.08741861979166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1753,
+      "loss/crossentropy": 2.2610700845718386,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22664773613214492,
+      "loss/reg": 0.0,
+      "step": 15400
+    },
+    {
+      "epoch": 0.10138157894736842,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.02672119140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1537,
+      "loss/crossentropy": 2.394396644830704,
+      "loss/hidden": 2.8859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25834341049194337,
+      "loss/reg": 0.0,
+      "step": 15410
+    },
+    {
+      "epoch": 0.10144736842105263,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.034895833333333334,
+      "learning_rate": 0.0001,
+      "loss": 3.1834,
+      "loss/crossentropy": 2.290709137916565,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2504901379346848,
+      "loss/reg": 0.0,
+      "step": 15420
+    },
+    {
+      "epoch": 0.10151315789473685,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 5.647508748372396,
+      "learning_rate": 0.0001,
+      "loss": 3.1705,
+      "loss/crossentropy": 2.597213554382324,
+      "loss/hidden": 2.9375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2691803678870201,
+      "loss/reg": 0.0,
+      "step": 15430
+    },
+    {
+      "epoch": 0.10157894736842105,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.07888997395833333,
+      "learning_rate": 0.0001,
+      "loss": 3.1069,
+      "loss/crossentropy": 2.1887213230133056,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2691598206758499,
+      "loss/reg": 0.0,
+      "step": 15440
+    },
+    {
+      "epoch": 0.10164473684210526,
+      "grad_norm": 3.109375,
+      "grad_norm_var": 0.0681793212890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1639,
+      "loss/crossentropy": 2.097235471010208,
+      "loss/hidden": 2.9625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2574038878083229,
+      "loss/reg": 0.0,
+      "step": 15450
+    },
+    {
+      "epoch": 0.10171052631578947,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.0678863525390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1301,
+      "loss/crossentropy": 2.2443934082984924,
+      "loss/hidden": 2.9046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24395571500062943,
+      "loss/reg": 0.0,
+      "step": 15460
+    },
+    {
+      "epoch": 0.10177631578947369,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.1054595947265625,
+      "learning_rate": 0.0001,
+      "loss": 3.2604,
+      "loss/crossentropy": 2.429443156719208,
+      "loss/hidden": 2.990625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2746900498867035,
+      "loss/reg": 0.0,
+      "step": 15470
+    },
+    {
+      "epoch": 0.1018421052631579,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.1326080322265625,
+      "learning_rate": 0.0001,
+      "loss": 3.221,
+      "loss/crossentropy": 2.423279583454132,
+      "loss/hidden": 2.9359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27492027878761294,
+      "loss/reg": 0.0,
+      "step": 15480
+    },
+    {
+      "epoch": 0.1019078947368421,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.1138092041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.257,
+      "loss/crossentropy": 2.648502016067505,
+      "loss/hidden": 2.99375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2935266062617302,
+      "loss/reg": 0.0,
+      "step": 15490
+    },
+    {
+      "epoch": 0.10197368421052631,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.1727203369140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1197,
+      "loss/crossentropy": 2.485008704662323,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2535603493452072,
+      "loss/reg": 0.0,
+      "step": 15500
+    },
+    {
+      "epoch": 0.10203947368421053,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.18816731770833334,
+      "learning_rate": 0.0001,
+      "loss": 3.1804,
+      "loss/crossentropy": 2.180854117870331,
+      "loss/hidden": 2.9703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24670835435390473,
+      "loss/reg": 0.0,
+      "step": 15510
+    },
+    {
+      "epoch": 0.10210526315789474,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.07157796223958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1547,
+      "loss/crossentropy": 2.5997716546058656,
+      "loss/hidden": 2.9765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2878054201602936,
+      "loss/reg": 0.0,
+      "step": 15520
+    },
+    {
+      "epoch": 0.10217105263157895,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.154443359375,
+      "learning_rate": 0.0001,
+      "loss": 3.1694,
+      "loss/crossentropy": 2.3764194369316103,
+      "loss/hidden": 2.8140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24656722396612168,
+      "loss/reg": 0.0,
+      "step": 15530
+    },
+    {
+      "epoch": 0.10223684210526315,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.25536702473958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1474,
+      "loss/crossentropy": 1.9735815048217773,
+      "loss/hidden": 2.8859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24598835706710814,
+      "loss/reg": 0.0,
+      "step": 15540
+    },
+    {
+      "epoch": 0.10230263157894737,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.1596832275390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1522,
+      "loss/crossentropy": 2.525629758834839,
+      "loss/hidden": 2.9921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33703051060438155,
+      "loss/reg": 0.0,
+      "step": 15550
+    },
+    {
+      "epoch": 0.10236842105263158,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.1472808837890625,
+      "learning_rate": 0.0001,
+      "loss": 3.212,
+      "loss/crossentropy": 2.03394900560379,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24415955394506456,
+      "loss/reg": 0.0,
+      "step": 15560
+    },
+    {
+      "epoch": 0.1024342105263158,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.07866923014322917,
+      "learning_rate": 0.0001,
+      "loss": 3.122,
+      "loss/crossentropy": 2.0943363308906555,
+      "loss/hidden": 2.9375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23014722466468812,
+      "loss/reg": 0.0,
+      "step": 15570
+    },
+    {
+      "epoch": 0.1025,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.057389322916666666,
+      "learning_rate": 0.0001,
+      "loss": 3.1268,
+      "loss/crossentropy": 2.033569025993347,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2175581559538841,
+      "loss/reg": 0.0,
+      "step": 15580
+    },
+    {
+      "epoch": 0.10256578947368421,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.08776041666666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0706,
+      "loss/crossentropy": 2.6018026471138,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2438505232334137,
+      "loss/reg": 0.0,
+      "step": 15590
+    },
+    {
+      "epoch": 0.10263157894736842,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.1537994384765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1439,
+      "loss/crossentropy": 2.196335256099701,
+      "loss/hidden": 3.021875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2767083361744881,
+      "loss/reg": 0.0,
+      "step": 15600
+    },
+    {
+      "epoch": 0.10269736842105263,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.07605692545572916,
+      "learning_rate": 0.0001,
+      "loss": 3.1523,
+      "loss/crossentropy": 2.326652777194977,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2691790975630283,
+      "loss/reg": 0.0,
+      "step": 15610
+    },
+    {
+      "epoch": 0.10276315789473685,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.09851455688476562,
+      "learning_rate": 0.0001,
+      "loss": 3.1315,
+      "loss/crossentropy": 2.219775491952896,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2539939820766449,
+      "loss/reg": 0.0,
+      "step": 15620
+    },
+    {
+      "epoch": 0.10282894736842105,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.14008560180664062,
+      "learning_rate": 0.0001,
+      "loss": 3.1501,
+      "loss/crossentropy": 2.284542143344879,
+      "loss/hidden": 2.9671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27733145356178285,
+      "loss/reg": 0.0,
+      "step": 15630
+    },
+    {
+      "epoch": 0.10289473684210526,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.10869115193684896,
+      "learning_rate": 0.0001,
+      "loss": 3.1621,
+      "loss/crossentropy": 2.4228519797325134,
+      "loss/hidden": 3.08125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32454578429460523,
+      "loss/reg": 0.0,
+      "step": 15640
+    },
+    {
+      "epoch": 0.10296052631578947,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.11193211873372395,
+      "learning_rate": 0.0001,
+      "loss": 3.1631,
+      "loss/crossentropy": 2.3004646062850953,
+      "loss/hidden": 2.9875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2865068309009075,
+      "loss/reg": 0.0,
+      "step": 15650
+    },
+    {
+      "epoch": 0.10302631578947369,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.08788655598958334,
+      "learning_rate": 0.0001,
+      "loss": 3.2413,
+      "loss/crossentropy": 2.3326223611831667,
+      "loss/hidden": 3.0703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2786666050553322,
+      "loss/reg": 0.0,
+      "step": 15660
+    },
+    {
+      "epoch": 0.1030921052631579,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.05933329264322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0799,
+      "loss/crossentropy": 2.308107304573059,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23495433181524278,
+      "loss/reg": 0.0,
+      "step": 15670
+    },
+    {
+      "epoch": 0.1031578947368421,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.04383926391601563,
+      "learning_rate": 0.0001,
+      "loss": 3.0889,
+      "loss/crossentropy": 2.3051168084144593,
+      "loss/hidden": 2.965625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.278233802318573,
+      "loss/reg": 0.0,
+      "step": 15680
+    },
+    {
+      "epoch": 0.10322368421052631,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 1.4827247619628907,
+      "learning_rate": 0.0001,
+      "loss": 3.1864,
+      "loss/crossentropy": 2.2644376397132873,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22414357215166092,
+      "loss/reg": 0.0,
+      "step": 15690
+    },
+    {
+      "epoch": 0.10328947368421053,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.962451171875,
+      "learning_rate": 0.0001,
+      "loss": 3.1546,
+      "loss/crossentropy": 2.424470007419586,
+      "loss/hidden": 2.8453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24368802309036255,
+      "loss/reg": 0.0,
+      "step": 15700
+    },
+    {
+      "epoch": 0.10335526315789474,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.07737223307291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1624,
+      "loss/crossentropy": 2.459938275814056,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23296410143375396,
+      "loss/reg": 0.0,
+      "step": 15710
+    },
+    {
+      "epoch": 0.10342105263157894,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.06443583170572917,
+      "learning_rate": 0.0001,
+      "loss": 3.1932,
+      "loss/crossentropy": 2.2153470873832704,
+      "loss/hidden": 3.146875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33586266040802004,
+      "loss/reg": 0.0,
+      "step": 15720
+    },
+    {
+      "epoch": 0.10348684210526315,
+      "grad_norm": 3.53125,
+      "grad_norm_var": 0.17069905598958332,
+      "learning_rate": 0.0001,
+      "loss": 3.257,
+      "loss/crossentropy": 2.2473284363746644,
+      "loss/hidden": 2.9140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24566184133291244,
+      "loss/reg": 0.0,
+      "step": 15730
+    },
+    {
+      "epoch": 0.10355263157894737,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.163916015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0707,
+      "loss/crossentropy": 2.620988368988037,
+      "loss/hidden": 3.0109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29941551238298414,
+      "loss/reg": 0.0,
+      "step": 15740
+    },
+    {
+      "epoch": 0.10361842105263158,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.35409749348958336,
+      "learning_rate": 0.0001,
+      "loss": 3.2423,
+      "loss/crossentropy": 2.3977234601974486,
+      "loss/hidden": 2.9765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26520802527666093,
+      "loss/reg": 0.0,
+      "step": 15750
+    },
+    {
+      "epoch": 0.1036842105263158,
+      "grad_norm": 2.375,
+      "grad_norm_var": 3.301877391050758e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2653,
+      "loss/crossentropy": 2.3854068875312806,
+      "loss/hidden": 3.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3163578942418098,
+      "loss/reg": 0.0,
+      "step": 15760
+    },
+    {
+      "epoch": 0.10375,
+      "grad_norm": 3.09375,
+      "grad_norm_var": 3.301877391679248e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1843,
+      "loss/crossentropy": 2.2795175433158876,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26798346936702727,
+      "loss/reg": 0.0,
+      "step": 15770
+    },
+    {
+      "epoch": 0.10381578947368421,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.1086334228515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1522,
+      "loss/crossentropy": 2.125895881652832,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23809853047132493,
+      "loss/reg": 0.0,
+      "step": 15780
+    },
+    {
+      "epoch": 0.10388157894736842,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.09806289672851562,
+      "learning_rate": 0.0001,
+      "loss": 3.121,
+      "loss/crossentropy": 2.3405375838279725,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23903141915798187,
+      "loss/reg": 0.0,
+      "step": 15790
+    },
+    {
+      "epoch": 0.10394736842105264,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.05409927368164062,
+      "learning_rate": 0.0001,
+      "loss": 3.1165,
+      "loss/crossentropy": 2.371259605884552,
+      "loss/hidden": 2.9640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24798257499933243,
+      "loss/reg": 0.0,
+      "step": 15800
+    },
+    {
+      "epoch": 0.10401315789473685,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.046727498372395836,
+      "learning_rate": 0.0001,
+      "loss": 3.0808,
+      "loss/crossentropy": 2.5483869433403017,
+      "loss/hidden": 3.0,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2987742185592651,
+      "loss/reg": 0.0,
+      "step": 15810
+    },
+    {
+      "epoch": 0.10407894736842105,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.43404541015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1913,
+      "loss/crossentropy": 2.3843488097190857,
+      "loss/hidden": 3.0671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.41426307857036593,
+      "loss/reg": 0.0,
+      "step": 15820
+    },
+    {
+      "epoch": 0.10414473684210526,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.42040913899739585,
+      "learning_rate": 0.0001,
+      "loss": 3.1412,
+      "loss/crossentropy": 2.451664757728577,
+      "loss/hidden": 3.0796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2949830338358879,
+      "loss/reg": 0.0,
+      "step": 15830
+    },
+    {
+      "epoch": 0.10421052631578948,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 12064558770995.855,
+      "learning_rate": 0.0001,
+      "loss": 3.2876,
+      "loss/crossentropy": 2.540658712387085,
+      "loss/hidden": 2.7984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3069792494177818,
+      "loss/reg": 0.0,
+      "step": 15840
+    },
+    {
+      "epoch": 0.10427631578947369,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.26617431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1409,
+      "loss/crossentropy": 2.3011590003967286,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23567215949296952,
+      "loss/reg": 0.0,
+      "step": 15850
+    },
+    {
+      "epoch": 0.10434210526315789,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.0716217041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.2032,
+      "loss/crossentropy": 2.624694299697876,
+      "loss/hidden": 2.91875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2735205709934235,
+      "loss/reg": 0.0,
+      "step": 15860
+    },
+    {
+      "epoch": 0.1044078947368421,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.037018839518229166,
+      "learning_rate": 0.0001,
+      "loss": 3.1547,
+      "loss/crossentropy": 2.3426929831504824,
+      "loss/hidden": 2.9796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26750322580337527,
+      "loss/reg": 0.0,
+      "step": 15870
+    },
+    {
+      "epoch": 0.10447368421052632,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.0340728759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1257,
+      "loss/crossentropy": 2.0538764238357543,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23220817893743514,
+      "loss/reg": 0.0,
+      "step": 15880
+    },
+    {
+      "epoch": 0.10453947368421053,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.10313084920247396,
+      "learning_rate": 0.0001,
+      "loss": 3.0791,
+      "loss/crossentropy": 2.4493046522140505,
+      "loss/hidden": 2.9375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.264581099152565,
+      "loss/reg": 0.0,
+      "step": 15890
+    },
+    {
+      "epoch": 0.10460526315789474,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.08592910766601562,
+      "learning_rate": 0.0001,
+      "loss": 3.1153,
+      "loss/crossentropy": 2.384776270389557,
+      "loss/hidden": 2.6609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20801746100187302,
+      "loss/reg": 0.0,
+      "step": 15900
+    },
+    {
+      "epoch": 0.10467105263157894,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.03535054524739583,
+      "learning_rate": 0.0001,
+      "loss": 3.1818,
+      "loss/crossentropy": 2.2780667304992677,
+      "loss/hidden": 3.059375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33164310455322266,
+      "loss/reg": 0.0,
+      "step": 15910
+    },
+    {
+      "epoch": 0.10473684210526316,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.04937744140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1317,
+      "loss/crossentropy": 2.3224631786346435,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26515267193317416,
+      "loss/reg": 0.0,
+      "step": 15920
+    },
+    {
+      "epoch": 0.10480263157894737,
+      "grad_norm": 3.375,
+      "grad_norm_var": 0.09521077473958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1393,
+      "loss/crossentropy": 2.26437486410141,
+      "loss/hidden": 2.890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25974428951740264,
+      "loss/reg": 0.0,
+      "step": 15930
+    },
+    {
+      "epoch": 0.10486842105263158,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.10048828125,
+      "learning_rate": 0.0001,
+      "loss": 3.203,
+      "loss/crossentropy": 2.216641104221344,
+      "loss/hidden": 2.9703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3172867178916931,
+      "loss/reg": 0.0,
+      "step": 15940
+    },
+    {
+      "epoch": 0.10493421052631578,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.0614898681640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0768,
+      "loss/crossentropy": 2.6155009508132934,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2578451469540596,
+      "loss/reg": 0.0,
+      "step": 15950
+    },
+    {
+      "epoch": 0.105,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.060269927978515624,
+      "learning_rate": 0.0001,
+      "loss": 3.161,
+      "loss/crossentropy": 2.5140405654907227,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2738805189728737,
+      "loss/reg": 0.0,
+      "step": 15960
+    },
+    {
+      "epoch": 0.10506578947368421,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.050176747639973956,
+      "learning_rate": 0.0001,
+      "loss": 3.0829,
+      "loss/crossentropy": 2.133234918117523,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2213057592511177,
+      "loss/reg": 0.0,
+      "step": 15970
+    },
+    {
+      "epoch": 0.10513157894736842,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.18106180826822918,
+      "learning_rate": 0.0001,
+      "loss": 3.2666,
+      "loss/crossentropy": 2.489140582084656,
+      "loss/hidden": 2.9421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2818035438656807,
+      "loss/reg": 0.0,
+      "step": 15980
+    },
+    {
+      "epoch": 0.10519736842105264,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.6164947509765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1722,
+      "loss/crossentropy": 2.51891051530838,
+      "loss/hidden": 2.7453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2563257798552513,
+      "loss/reg": 0.0,
+      "step": 15990
+    },
+    {
+      "epoch": 0.10526315789473684,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.8845987955729167,
+      "learning_rate": 0.0001,
+      "loss": 3.1763,
+      "loss/crossentropy": 2.2177582025527953,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2367611363530159,
+      "loss/reg": 0.0,
+      "step": 16000
+    },
+    {
+      "epoch": 0.10532894736842105,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.451318359375,
+      "learning_rate": 0.0001,
+      "loss": 3.1408,
+      "loss/crossentropy": 2.2017379879951475,
+      "loss/hidden": 2.7234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20062217488884926,
+      "loss/reg": 0.0,
+      "step": 16010
+    },
+    {
+      "epoch": 0.10539473684210526,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.050439453125,
+      "learning_rate": 0.0001,
+      "loss": 3.1466,
+      "loss/crossentropy": 2.2292946934700013,
+      "loss/hidden": 2.8625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2783455640077591,
+      "loss/reg": 0.0,
+      "step": 16020
+    },
+    {
+      "epoch": 0.10546052631578948,
+      "grad_norm": 1.984375,
+      "grad_norm_var": 0.1126129150390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1018,
+      "loss/crossentropy": 2.530960404872894,
+      "loss/hidden": 2.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23581028431653978,
+      "loss/reg": 0.0,
+      "step": 16030
+    },
+    {
+      "epoch": 0.10552631578947369,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.08625386555989584,
+      "learning_rate": 0.0001,
+      "loss": 3.1222,
+      "loss/crossentropy": 2.432818961143494,
+      "loss/hidden": 2.8390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23912513256072998,
+      "loss/reg": 0.0,
+      "step": 16040
+    },
+    {
+      "epoch": 0.10559210526315789,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.1727203369140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1108,
+      "loss/crossentropy": 2.481075167655945,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2601512670516968,
+      "loss/reg": 0.0,
+      "step": 16050
+    },
+    {
+      "epoch": 0.1056578947368421,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.22974853515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1102,
+      "loss/crossentropy": 2.347836995124817,
+      "loss/hidden": 2.9515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2967532381415367,
+      "loss/reg": 0.0,
+      "step": 16060
+    },
+    {
+      "epoch": 0.10572368421052632,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 3.920637003580729,
+      "learning_rate": 0.0001,
+      "loss": 3.1789,
+      "loss/crossentropy": 2.2461806178092956,
+      "loss/hidden": 3.0078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2796615108847618,
+      "loss/reg": 0.0,
+      "step": 16070
+    },
+    {
+      "epoch": 0.10578947368421053,
+      "grad_norm": 3.234375,
+      "grad_norm_var": 3.8184529622395833,
+      "learning_rate": 0.0001,
+      "loss": 3.2028,
+      "loss/crossentropy": 2.2602067947387696,
+      "loss/hidden": 3.059375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27418701648712157,
+      "loss/reg": 0.0,
+      "step": 16080
+    },
+    {
+      "epoch": 0.10585526315789473,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.0978912353515625,
+      "learning_rate": 0.0001,
+      "loss": 3.087,
+      "loss/crossentropy": 2.4076030969619753,
+      "loss/hidden": 2.765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2339254654943943,
+      "loss/reg": 0.0,
+      "step": 16090
+    },
+    {
+      "epoch": 0.10592105263157894,
+      "grad_norm": 8.8125,
+      "grad_norm_var": 2.634016927083333,
+      "learning_rate": 0.0001,
+      "loss": 3.2404,
+      "loss/crossentropy": 1.9294680893421172,
+      "loss/hidden": 2.8984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24796108528971672,
+      "loss/reg": 0.0,
+      "step": 16100
+    },
+    {
+      "epoch": 0.10598684210526316,
+      "grad_norm": 2.5,
+      "grad_norm_var": 2.7831013997395835,
+      "learning_rate": 0.0001,
+      "loss": 3.1866,
+      "loss/crossentropy": 1.8769205152988433,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2611444815993309,
+      "loss/reg": 0.0,
+      "step": 16110
+    },
+    {
+      "epoch": 0.10605263157894737,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.76617431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.2357,
+      "loss/crossentropy": 2.2888787627220153,
+      "loss/hidden": 2.90625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2839828670024872,
+      "loss/reg": 0.0,
+      "step": 16120
+    },
+    {
+      "epoch": 0.10611842105263158,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.04062093098958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1608,
+      "loss/crossentropy": 2.313342797756195,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2872613161802292,
+      "loss/reg": 0.0,
+      "step": 16130
+    },
+    {
+      "epoch": 0.10618421052631578,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.4191691080729167,
+      "learning_rate": 0.0001,
+      "loss": 3.1753,
+      "loss/crossentropy": 2.6026643037796022,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28212335854768755,
+      "loss/reg": 0.0,
+      "step": 16140
+    },
+    {
+      "epoch": 0.10625,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.41529541015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1346,
+      "loss/crossentropy": 2.556915271282196,
+      "loss/hidden": 3.1078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2837150752544403,
+      "loss/reg": 0.0,
+      "step": 16150
+    },
+    {
+      "epoch": 0.10631578947368421,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.053441365559895836,
+      "learning_rate": 0.0001,
+      "loss": 3.1471,
+      "loss/crossentropy": 2.399764931201935,
+      "loss/hidden": 2.8265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2460502192378044,
+      "loss/reg": 0.0,
+      "step": 16160
+    },
+    {
+      "epoch": 0.10638157894736842,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.8765462239583334,
+      "learning_rate": 0.0001,
+      "loss": 3.2831,
+      "loss/crossentropy": 2.2979734420776365,
+      "loss/hidden": 3.06875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2876197725534439,
+      "loss/reg": 0.0,
+      "step": 16170
+    },
+    {
+      "epoch": 0.10644736842105264,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.9296834309895833,
+      "learning_rate": 0.0001,
+      "loss": 3.2187,
+      "loss/crossentropy": 2.2085362553596495,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24199980795383452,
+      "loss/reg": 0.0,
+      "step": 16180
+    },
+    {
+      "epoch": 0.10651315789473684,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.06541239420572917,
+      "learning_rate": 0.0001,
+      "loss": 3.1524,
+      "loss/crossentropy": 2.4575427293777468,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2581774353981018,
+      "loss/reg": 0.0,
+      "step": 16190
+    },
+    {
+      "epoch": 0.10657894736842105,
+      "grad_norm": 3.34375,
+      "grad_norm_var": 0.2840983072916667,
+      "learning_rate": 0.0001,
+      "loss": 3.2356,
+      "loss/crossentropy": 2.444900369644165,
+      "loss/hidden": 3.1625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.295256008207798,
+      "loss/reg": 0.0,
+      "step": 16200
+    },
+    {
+      "epoch": 0.10664473684210526,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.27533137003580727,
+      "learning_rate": 0.0001,
+      "loss": 3.0994,
+      "loss/crossentropy": 2.5305007696151733,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2888329938054085,
+      "loss/reg": 0.0,
+      "step": 16210
+    },
+    {
+      "epoch": 0.10671052631578948,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.6374224344889323,
+      "learning_rate": 0.0001,
+      "loss": 3.1837,
+      "loss/crossentropy": 2.394989788532257,
+      "loss/hidden": 2.790625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22691280096769334,
+      "loss/reg": 0.0,
+      "step": 16220
+    },
+    {
+      "epoch": 0.10677631578947368,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.6025950113932291,
+      "learning_rate": 0.0001,
+      "loss": 3.1035,
+      "loss/crossentropy": 2.3761568784713747,
+      "loss/hidden": 2.9078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25212556272745135,
+      "loss/reg": 0.0,
+      "step": 16230
+    },
+    {
+      "epoch": 0.10684210526315789,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.07968343098958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1507,
+      "loss/crossentropy": 2.1518751621246337,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25603573620319364,
+      "loss/reg": 0.0,
+      "step": 16240
+    },
+    {
+      "epoch": 0.1069078947368421,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.0736968994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0942,
+      "loss/crossentropy": 2.2170523405075073,
+      "loss/hidden": 2.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2497616469860077,
+      "loss/reg": 0.0,
+      "step": 16250
+    },
+    {
+      "epoch": 0.10697368421052632,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.15055338541666666,
+      "learning_rate": 0.0001,
+      "loss": 3.18,
+      "loss/crossentropy": 2.2126652002334595,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2371530830860138,
+      "loss/reg": 0.0,
+      "step": 16260
+    },
+    {
+      "epoch": 0.10703947368421053,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.1051666259765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1175,
+      "loss/crossentropy": 2.5849027037620544,
+      "loss/hidden": 3.00625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31244791448116305,
+      "loss/reg": 0.0,
+      "step": 16270
+    },
+    {
+      "epoch": 0.10710526315789473,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.05666910807291667,
+      "learning_rate": 0.0001,
+      "loss": 3.2064,
+      "loss/crossentropy": 1.9444570660591125,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2180204689502716,
+      "loss/reg": 0.0,
+      "step": 16280
+    },
+    {
+      "epoch": 0.10717105263157894,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.08806050618489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0722,
+      "loss/crossentropy": 2.3104967713356017,
+      "loss/hidden": 2.9296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2631675943732262,
+      "loss/reg": 0.0,
+      "step": 16290
+    },
+    {
+      "epoch": 0.10723684210526316,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.018146769205729166,
+      "learning_rate": 0.0001,
+      "loss": 3.0587,
+      "loss/crossentropy": 2.1140322208404543,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2300342008471489,
+      "loss/reg": 0.0,
+      "step": 16300
+    },
+    {
+      "epoch": 0.10730263157894737,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.036774698893229166,
+      "learning_rate": 0.0001,
+      "loss": 3.2022,
+      "loss/crossentropy": 2.5614047765731813,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2788904532790184,
+      "loss/reg": 0.0,
+      "step": 16310
+    },
+    {
+      "epoch": 0.10736842105263159,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.21725972493489584,
+      "learning_rate": 0.0001,
+      "loss": 3.1511,
+      "loss/crossentropy": 1.8924081802368165,
+      "loss/hidden": 2.8765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2040714904665947,
+      "loss/reg": 0.0,
+      "step": 16320
+    },
+    {
+      "epoch": 0.10743421052631578,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.23220926920572918,
+      "learning_rate": 0.0001,
+      "loss": 3.1014,
+      "loss/crossentropy": 2.342508816719055,
+      "loss/hidden": 2.884375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27500451505184176,
+      "loss/reg": 0.0,
+      "step": 16330
+    },
+    {
+      "epoch": 0.1075,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.5662424723307292,
+      "learning_rate": 0.0001,
+      "loss": 3.2912,
+      "loss/crossentropy": 2.3959147095680238,
+      "loss/hidden": 2.9625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2863708436489105,
+      "loss/reg": 0.0,
+      "step": 16340
+    },
+    {
+      "epoch": 0.10756578947368421,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.11278889973958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1532,
+      "loss/crossentropy": 2.166390228271484,
+      "loss/hidden": 2.86875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23770884573459625,
+      "loss/reg": 0.0,
+      "step": 16350
+    },
+    {
+      "epoch": 0.10763157894736843,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.06780776977539063,
+      "learning_rate": 0.0001,
+      "loss": 3.0399,
+      "loss/crossentropy": 2.3057939767837525,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2797847852110863,
+      "loss/reg": 0.0,
+      "step": 16360
+    },
+    {
+      "epoch": 0.10769736842105262,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.038826243082682295,
+      "learning_rate": 0.0001,
+      "loss": 3.12,
+      "loss/crossentropy": 2.265635335445404,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23047662824392318,
+      "loss/reg": 0.0,
+      "step": 16370
+    },
+    {
+      "epoch": 0.10776315789473684,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.08754781087239584,
+      "learning_rate": 0.0001,
+      "loss": 3.0758,
+      "loss/crossentropy": 2.6614980459213258,
+      "loss/hidden": 2.6609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23254811465740205,
+      "loss/reg": 0.0,
+      "step": 16380
+    },
+    {
+      "epoch": 0.10782894736842105,
+      "grad_norm": 3.390625,
+      "grad_norm_var": 0.14463882446289061,
+      "learning_rate": 0.0001,
+      "loss": 3.1805,
+      "loss/crossentropy": 2.419999623298645,
+      "loss/hidden": 3.0984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3057109400629997,
+      "loss/reg": 0.0,
+      "step": 16390
+    },
+    {
+      "epoch": 0.10789473684210527,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.12823893229166666,
+      "learning_rate": 0.0001,
+      "loss": 3.1467,
+      "loss/crossentropy": 2.145538020133972,
+      "loss/hidden": 2.8890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23782579749822616,
+      "loss/reg": 0.0,
+      "step": 16400
+    },
+    {
+      "epoch": 0.10796052631578948,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.07416890462239584,
+      "learning_rate": 0.0001,
+      "loss": 3.1471,
+      "loss/crossentropy": 2.3262511491775513,
+      "loss/hidden": 2.7125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23651919960975648,
+      "loss/reg": 0.0,
+      "step": 16410
+    },
+    {
+      "epoch": 0.10802631578947368,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.052245076497395834,
+      "learning_rate": 0.0001,
+      "loss": 3.1573,
+      "loss/crossentropy": 2.446836495399475,
+      "loss/hidden": 2.6953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22134366482496262,
+      "loss/reg": 0.0,
+      "step": 16420
+    },
+    {
+      "epoch": 0.10809210526315789,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.12043355305989584,
+      "learning_rate": 0.0001,
+      "loss": 3.2359,
+      "loss/crossentropy": 2.3113945603370665,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2689176544547081,
+      "loss/reg": 0.0,
+      "step": 16430
+    },
+    {
+      "epoch": 0.1081578947368421,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.10204671223958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1679,
+      "loss/crossentropy": 2.422752869129181,
+      "loss/hidden": 2.9703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.306715852022171,
+      "loss/reg": 0.0,
+      "step": 16440
+    },
+    {
+      "epoch": 0.10822368421052632,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.06880594889322916,
+      "learning_rate": 0.0001,
+      "loss": 3.147,
+      "loss/crossentropy": 2.048931634426117,
+      "loss/hidden": 2.93125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2303798720240593,
+      "loss/reg": 0.0,
+      "step": 16450
+    },
+    {
+      "epoch": 0.10828947368421053,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.15419514973958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1877,
+      "loss/crossentropy": 2.68409343957901,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25028681606054304,
+      "loss/reg": 0.0,
+      "step": 16460
+    },
+    {
+      "epoch": 0.10835526315789473,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.09561258951822917,
+      "learning_rate": 0.0001,
+      "loss": 3.0881,
+      "loss/crossentropy": 2.3087923645973207,
+      "loss/hidden": 2.9171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2732239991426468,
+      "loss/reg": 0.0,
+      "step": 16470
+    },
+    {
+      "epoch": 0.10842105263157895,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.14474283854166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1675,
+      "loss/crossentropy": 2.3147490501403807,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25520786345005037,
+      "loss/reg": 0.0,
+      "step": 16480
+    },
+    {
+      "epoch": 0.10848684210526316,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.12763671875,
+      "learning_rate": 0.0001,
+      "loss": 3.1948,
+      "loss/crossentropy": 2.4046076416969298,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24635857343673706,
+      "loss/reg": 0.0,
+      "step": 16490
+    },
+    {
+      "epoch": 0.10855263157894737,
+      "grad_norm": 3.234375,
+      "grad_norm_var": 0.19675267537434896,
+      "learning_rate": 0.0001,
+      "loss": 3.134,
+      "loss/crossentropy": 2.1358281135559083,
+      "loss/hidden": 2.684375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20830129384994506,
+      "loss/reg": 0.0,
+      "step": 16500
+    },
+    {
+      "epoch": 0.10861842105263157,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.20981216430664062,
+      "learning_rate": 0.0001,
+      "loss": 3.1161,
+      "loss/crossentropy": 2.2100176930427553,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22965313643217086,
+      "loss/reg": 0.0,
+      "step": 16510
+    },
+    {
+      "epoch": 0.10868421052631579,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.025813802083333334,
+      "learning_rate": 0.0001,
+      "loss": 3.1402,
+      "loss/crossentropy": 2.1080865144729612,
+      "loss/hidden": 2.9640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24372481554746628,
+      "loss/reg": 0.0,
+      "step": 16520
+    },
+    {
+      "epoch": 0.10875,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.03843994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.112,
+      "loss/crossentropy": 2.184442663192749,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25090422928333284,
+      "loss/reg": 0.0,
+      "step": 16530
+    },
+    {
+      "epoch": 0.10881578947368421,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.03850682576497396,
+      "learning_rate": 0.0001,
+      "loss": 3.0917,
+      "loss/crossentropy": 2.2426281213760375,
+      "loss/hidden": 2.9359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26249563246965407,
+      "loss/reg": 0.0,
+      "step": 16540
+    },
+    {
+      "epoch": 0.10888157894736843,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.037082672119140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1533,
+      "loss/crossentropy": 2.4359527707099913,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28122682869434357,
+      "loss/reg": 0.0,
+      "step": 16550
+    },
+    {
+      "epoch": 0.10894736842105263,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.06422526041666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1086,
+      "loss/crossentropy": 2.679885816574097,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25820834636688234,
+      "loss/reg": 0.0,
+      "step": 16560
+    },
+    {
+      "epoch": 0.10901315789473684,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.033543904622395836,
+      "learning_rate": 0.0001,
+      "loss": 3.1226,
+      "loss/crossentropy": 2.230025511980057,
+      "loss/hidden": 2.8640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24839217215776443,
+      "loss/reg": 0.0,
+      "step": 16570
+    },
+    {
+      "epoch": 0.10907894736842105,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.08507486979166666,
+      "learning_rate": 0.0001,
+      "loss": 3.158,
+      "loss/crossentropy": 2.3745269417762755,
+      "loss/hidden": 2.740625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22690069079399108,
+      "loss/reg": 0.0,
+      "step": 16580
+    },
+    {
+      "epoch": 0.10914473684210527,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.07500712076822917,
+      "learning_rate": 0.0001,
+      "loss": 3.0464,
+      "loss/crossentropy": 2.3726358652114867,
+      "loss/hidden": 2.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22285908311605454,
+      "loss/reg": 0.0,
+      "step": 16590
+    },
+    {
+      "epoch": 0.10921052631578948,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 0.0905426025390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0802,
+      "loss/crossentropy": 2.297460901737213,
+      "loss/hidden": 2.9359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24248487651348113,
+      "loss/reg": 0.0,
+      "step": 16600
+    },
+    {
+      "epoch": 0.10927631578947368,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.42377827962239584,
+      "learning_rate": 0.0001,
+      "loss": 3.1369,
+      "loss/crossentropy": 2.1492306351661683,
+      "loss/hidden": 2.978125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25582308024168016,
+      "loss/reg": 0.0,
+      "step": 16610
+    },
+    {
+      "epoch": 0.1093421052631579,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.38531494140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1291,
+      "loss/crossentropy": 2.1944116175174715,
+      "loss/hidden": 3.025,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26006165742874143,
+      "loss/reg": 0.0,
+      "step": 16620
+    },
+    {
+      "epoch": 0.1094078947368421,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.031305948893229164,
+      "learning_rate": 0.0001,
+      "loss": 3.0966,
+      "loss/crossentropy": 2.2580446600914,
+      "loss/hidden": 2.9140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2571749180555344,
+      "loss/reg": 0.0,
+      "step": 16630
+    },
+    {
+      "epoch": 0.10947368421052632,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.08385009765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0758,
+      "loss/crossentropy": 2.711108660697937,
+      "loss/hidden": 2.790625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23514284193515778,
+      "loss/reg": 0.0,
+      "step": 16640
+    },
+    {
+      "epoch": 0.10953947368421052,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.39968973795572915,
+      "learning_rate": 0.0001,
+      "loss": 3.212,
+      "loss/crossentropy": 2.318574833869934,
+      "loss/hidden": 2.953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3364941954612732,
+      "loss/reg": 0.0,
+      "step": 16650
+    },
+    {
+      "epoch": 0.10960526315789473,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.5502431233723958,
+      "learning_rate": 0.0001,
+      "loss": 3.1398,
+      "loss/crossentropy": 2.5402591228485107,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2698328003287315,
+      "loss/reg": 0.0,
+      "step": 16660
+    },
+    {
+      "epoch": 0.10967105263157895,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.7628651936848958,
+      "learning_rate": 0.0001,
+      "loss": 3.2125,
+      "loss/crossentropy": 2.4290089428424837,
+      "loss/hidden": 2.959375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23459196090698242,
+      "loss/reg": 0.0,
+      "step": 16670
+    },
+    {
+      "epoch": 0.10973684210526316,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.5333984375,
+      "learning_rate": 0.0001,
+      "loss": 3.1568,
+      "loss/crossentropy": 2.4715004682540895,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2527153715491295,
+      "loss/reg": 0.0,
+      "step": 16680
+    },
+    {
+      "epoch": 0.10980263157894737,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.01929931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1188,
+      "loss/crossentropy": 2.565124809741974,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2764893934130669,
+      "loss/reg": 0.0,
+      "step": 16690
+    },
+    {
+      "epoch": 0.10986842105263157,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.7143287658691406,
+      "learning_rate": 0.0001,
+      "loss": 3.1851,
+      "loss/crossentropy": 2.1823901176452636,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2804734021425247,
+      "loss/reg": 0.0,
+      "step": 16700
+    },
+    {
+      "epoch": 0.10993421052631579,
+      "grad_norm": 3.75,
+      "grad_norm_var": 0.800158437093099,
+      "learning_rate": 0.0001,
+      "loss": 3.1599,
+      "loss/crossentropy": 2.339951229095459,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23388542532920836,
+      "loss/reg": 0.0,
+      "step": 16710
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 3.0625,
+      "grad_norm_var": 0.46857808430989584,
+      "learning_rate": 0.0001,
+      "loss": 3.168,
+      "loss/crossentropy": 2.5235843658447266,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25846762359142306,
+      "loss/reg": 0.0,
+      "step": 16720
+    },
+    {
+      "epoch": 0.11006578947368421,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.2997792561848958,
+      "learning_rate": 0.0001,
+      "loss": 3.199,
+      "loss/crossentropy": 2.4483086824417115,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24846359938383103,
+      "loss/reg": 0.0,
+      "step": 16730
+    },
+    {
+      "epoch": 0.11013157894736843,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.0768218994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1155,
+      "loss/crossentropy": 2.318689703941345,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25379282981157303,
+      "loss/reg": 0.0,
+      "step": 16740
+    },
+    {
+      "epoch": 0.11019736842105263,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.04757258097330729,
+      "learning_rate": 0.0001,
+      "loss": 3.0428,
+      "loss/crossentropy": 2.3445772767066955,
+      "loss/hidden": 2.78125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26050496101379395,
+      "loss/reg": 0.0,
+      "step": 16750
+    },
+    {
+      "epoch": 0.11026315789473684,
+      "grad_norm": 3.265625,
+      "grad_norm_var": 0.09231669108072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1426,
+      "loss/crossentropy": 2.2978576898574827,
+      "loss/hidden": 2.821875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23892502784729003,
+      "loss/reg": 0.0,
+      "step": 16760
+    },
+    {
+      "epoch": 0.11032894736842105,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.09435221354166666,
+      "learning_rate": 0.0001,
+      "loss": 3.0687,
+      "loss/crossentropy": 2.15471470952034,
+      "loss/hidden": 2.828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22209101915359497,
+      "loss/reg": 0.0,
+      "step": 16770
+    },
+    {
+      "epoch": 0.11039473684210527,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.084326171875,
+      "learning_rate": 0.0001,
+      "loss": 3.0934,
+      "loss/crossentropy": 2.0944029092788696,
+      "loss/hidden": 2.9328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2732590340077877,
+      "loss/reg": 0.0,
+      "step": 16780
+    },
+    {
+      "epoch": 0.11046052631578947,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.07172749837239584,
+      "learning_rate": 0.0001,
+      "loss": 3.196,
+      "loss/crossentropy": 2.4396159648895264,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25805698037147523,
+      "loss/reg": 0.0,
+      "step": 16790
+    },
+    {
+      "epoch": 0.11052631578947368,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.04521382649739583,
+      "learning_rate": 0.0001,
+      "loss": 3.1492,
+      "loss/crossentropy": 2.3519849300384523,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24649366587400437,
+      "loss/reg": 0.0,
+      "step": 16800
+    },
+    {
+      "epoch": 0.1105921052631579,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.06435445149739584,
+      "learning_rate": 0.0001,
+      "loss": 3.1366,
+      "loss/crossentropy": 2.2999491453170777,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2394917294383049,
+      "loss/reg": 0.0,
+      "step": 16810
+    },
+    {
+      "epoch": 0.11065789473684211,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.09391988118489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1441,
+      "loss/crossentropy": 2.4930548071861267,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2843811124563217,
+      "loss/reg": 0.0,
+      "step": 16820
+    },
+    {
+      "epoch": 0.11072368421052632,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.23037007649739583,
+      "learning_rate": 0.0001,
+      "loss": 3.1744,
+      "loss/crossentropy": 2.327615487575531,
+      "loss/hidden": 3.0484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25857883393764497,
+      "loss/reg": 0.0,
+      "step": 16830
+    },
+    {
+      "epoch": 0.11078947368421052,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.08505859375,
+      "learning_rate": 0.0001,
+      "loss": 3.1966,
+      "loss/crossentropy": 2.0340147018432617,
+      "loss/hidden": 2.7625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22139777690172197,
+      "loss/reg": 0.0,
+      "step": 16840
+    },
+    {
+      "epoch": 0.11085526315789473,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.04459228515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0559,
+      "loss/crossentropy": 2.327996277809143,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2415407806634903,
+      "loss/reg": 0.0,
+      "step": 16850
+    },
+    {
+      "epoch": 0.11092105263157895,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.15735677083333333,
+      "learning_rate": 0.0001,
+      "loss": 3.1824,
+      "loss/crossentropy": 2.3824142098426817,
+      "loss/hidden": 3.0859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28444231003522874,
+      "loss/reg": 0.0,
+      "step": 16860
+    },
+    {
+      "epoch": 0.11098684210526316,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.2694498697916667,
+      "learning_rate": 0.0001,
+      "loss": 3.1079,
+      "loss/crossentropy": 2.4686101198196413,
+      "loss/hidden": 2.740625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23947114795446395,
+      "loss/reg": 0.0,
+      "step": 16870
+    },
+    {
+      "epoch": 0.11105263157894738,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.2833811442057292,
+      "learning_rate": 0.0001,
+      "loss": 3.0544,
+      "loss/crossentropy": 2.2384608387947083,
+      "loss/hidden": 3.23125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29755171537399294,
+      "loss/reg": 0.0,
+      "step": 16880
+    },
+    {
+      "epoch": 0.11111842105263157,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.11591389973958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1316,
+      "loss/crossentropy": 2.065184140205383,
+      "loss/hidden": 2.7421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22740320414304732,
+      "loss/reg": 0.0,
+      "step": 16890
+    },
+    {
+      "epoch": 0.11118421052631579,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.14423421223958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1867,
+      "loss/crossentropy": 2.241389238834381,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2434727743268013,
+      "loss/reg": 0.0,
+      "step": 16900
+    },
+    {
+      "epoch": 0.11125,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.14661458333333333,
+      "learning_rate": 0.0001,
+      "loss": 3.1063,
+      "loss/crossentropy": 2.1571604132652284,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21824948787689208,
+      "loss/reg": 0.0,
+      "step": 16910
+    },
+    {
+      "epoch": 0.11131578947368422,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.9939605712890625,
+      "learning_rate": 0.0001,
+      "loss": 3.2191,
+      "loss/crossentropy": 2.4117300748825072,
+      "loss/hidden": 3.2203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33291524201631545,
+      "loss/reg": 0.0,
+      "step": 16920
+    },
+    {
+      "epoch": 0.11138157894736841,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.9513580322265625,
+      "learning_rate": 0.0001,
+      "loss": 3.2001,
+      "loss/crossentropy": 2.490310883522034,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2631781131029129,
+      "loss/reg": 0.0,
+      "step": 16930
+    },
+    {
+      "epoch": 0.11144736842105263,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.12056884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.203,
+      "loss/crossentropy": 2.3007746815681456,
+      "loss/hidden": 2.9046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2449628993868828,
+      "loss/reg": 0.0,
+      "step": 16940
+    },
+    {
+      "epoch": 0.11151315789473684,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.11777242024739583,
+      "learning_rate": 0.0001,
+      "loss": 3.1245,
+      "loss/crossentropy": 2.3470576763153077,
+      "loss/hidden": 2.83125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24132359698414801,
+      "loss/reg": 0.0,
+      "step": 16950
+    },
+    {
+      "epoch": 0.11157894736842106,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.05214742024739583,
+      "learning_rate": 0.0001,
+      "loss": 3.115,
+      "loss/crossentropy": 2.4354201793670653,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22888378351926802,
+      "loss/reg": 0.0,
+      "step": 16960
+    },
+    {
+      "epoch": 0.11164473684210527,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.07139383951822917,
+      "learning_rate": 0.0001,
+      "loss": 3.1537,
+      "loss/crossentropy": 2.1988754749298094,
+      "loss/hidden": 2.7453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22949053347110748,
+      "loss/reg": 0.0,
+      "step": 16970
+    },
+    {
+      "epoch": 0.11171052631578947,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.06669921875,
+      "learning_rate": 0.0001,
+      "loss": 3.113,
+      "loss/crossentropy": 2.2436501502990724,
+      "loss/hidden": 2.9671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26571269184350965,
+      "loss/reg": 0.0,
+      "step": 16980
+    },
+    {
+      "epoch": 0.11177631578947368,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.1141998291015625,
+      "learning_rate": 0.0001,
+      "loss": 3.2373,
+      "loss/crossentropy": 2.0406009197235107,
+      "loss/hidden": 2.89375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2788648784160614,
+      "loss/reg": 0.0,
+      "step": 16990
+    },
+    {
+      "epoch": 0.1118421052631579,
+      "grad_norm": 3.5,
+      "grad_norm_var": 0.13699544270833333,
+      "learning_rate": 0.0001,
+      "loss": 3.1371,
+      "loss/crossentropy": 2.282533049583435,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27586452960968016,
+      "loss/reg": 0.0,
+      "step": 17000
+    },
+    {
+      "epoch": 0.11190789473684211,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.13681233723958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1898,
+      "loss/crossentropy": 2.192341995239258,
+      "loss/hidden": 2.9890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2728851273655891,
+      "loss/reg": 0.0,
+      "step": 17010
+    },
+    {
+      "epoch": 0.11197368421052632,
+      "grad_norm": 2.9375,
+      "grad_norm_var": 0.08626302083333333,
+      "learning_rate": 0.0001,
+      "loss": 3.1208,
+      "loss/crossentropy": 2.036851680278778,
+      "loss/hidden": 3.090625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29492041319608686,
+      "loss/reg": 0.0,
+      "step": 17020
+    },
+    {
+      "epoch": 0.11203947368421052,
+      "grad_norm": 3.171875,
+      "grad_norm_var": 0.38601888020833336,
+      "learning_rate": 0.0001,
+      "loss": 3.1351,
+      "loss/crossentropy": 2.013105309009552,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21977066546678542,
+      "loss/reg": 0.0,
+      "step": 17030
+    },
+    {
+      "epoch": 0.11210526315789474,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.515862782796224,
+      "learning_rate": 0.0001,
+      "loss": 3.0822,
+      "loss/crossentropy": 2.2523082733154296,
+      "loss/hidden": 2.8671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23816863894462587,
+      "loss/reg": 0.0,
+      "step": 17040
+    },
+    {
+      "epoch": 0.11217105263157895,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.13596598307291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0894,
+      "loss/crossentropy": 2.474255657196045,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28221793919801713,
+      "loss/reg": 0.0,
+      "step": 17050
+    },
+    {
+      "epoch": 0.11223684210526316,
+      "grad_norm": 2.375,
+      "grad_norm_var": 2.8374436804564963e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2437,
+      "loss/crossentropy": 2.5942065715789795,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26605169773101806,
+      "loss/reg": 0.0,
+      "step": 17060
+    },
+    {
+      "epoch": 0.11230263157894736,
+      "grad_norm": 1.984375,
+      "grad_norm_var": 0.05279541015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0525,
+      "loss/crossentropy": 2.034271013736725,
+      "loss/hidden": 2.8875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2154896892607212,
+      "loss/reg": 0.0,
+      "step": 17070
+    },
+    {
+      "epoch": 0.11236842105263158,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.050902303059895834,
+      "learning_rate": 0.0001,
+      "loss": 3.1033,
+      "loss/crossentropy": 2.229046678543091,
+      "loss/hidden": 2.7578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22233688235282897,
+      "loss/reg": 0.0,
+      "step": 17080
+    },
+    {
+      "epoch": 0.11243421052631579,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.04163309733072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1293,
+      "loss/crossentropy": 2.4799925684928894,
+      "loss/hidden": 3.0203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2588403090834618,
+      "loss/reg": 0.0,
+      "step": 17090
+    },
+    {
+      "epoch": 0.1125,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.1295074462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0953,
+      "loss/crossentropy": 2.308819645643234,
+      "loss/hidden": 2.7390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23669061064720154,
+      "loss/reg": 0.0,
+      "step": 17100
+    },
+    {
+      "epoch": 0.11256578947368422,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.7806142171223959,
+      "learning_rate": 0.0001,
+      "loss": 3.1353,
+      "loss/crossentropy": 2.3123676419258117,
+      "loss/hidden": 2.9859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29565141499042513,
+      "loss/reg": 0.0,
+      "step": 17110
+    },
+    {
+      "epoch": 0.11263157894736842,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.08242085774739584,
+      "learning_rate": 0.0001,
+      "loss": 3.0662,
+      "loss/crossentropy": 2.262893891334534,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23198095709085464,
+      "loss/reg": 0.0,
+      "step": 17120
+    },
+    {
+      "epoch": 0.11269736842105263,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.1591949462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0855,
+      "loss/crossentropy": 2.1816434502601623,
+      "loss/hidden": 3.01875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2444119155406952,
+      "loss/reg": 0.0,
+      "step": 17130
+    },
+    {
+      "epoch": 0.11276315789473684,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.14752197265625,
+      "learning_rate": 0.0001,
+      "loss": 3.1414,
+      "loss/crossentropy": 2.4847304582595826,
+      "loss/hidden": 2.8203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26482110619544985,
+      "loss/reg": 0.0,
+      "step": 17140
+    },
+    {
+      "epoch": 0.11282894736842106,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.09546610514322916,
+      "learning_rate": 0.0001,
+      "loss": 3.0923,
+      "loss/crossentropy": 2.5822364687919617,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26294154226779937,
+      "loss/reg": 0.0,
+      "step": 17150
+    },
+    {
+      "epoch": 0.11289473684210527,
+      "grad_norm": 3.796875,
+      "grad_norm_var": 0.19820963541666667,
+      "learning_rate": 0.0001,
+      "loss": 3.0656,
+      "loss/crossentropy": 2.3146368622779847,
+      "loss/hidden": 2.9109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23589982092380524,
+      "loss/reg": 0.0,
+      "step": 17160
+    },
+    {
+      "epoch": 0.11296052631578947,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 1.0957618713378907,
+      "learning_rate": 0.0001,
+      "loss": 3.1693,
+      "loss/crossentropy": 2.355622184276581,
+      "loss/hidden": 2.9609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2783960849046707,
+      "loss/reg": 0.0,
+      "step": 17170
+    },
+    {
+      "epoch": 0.11302631578947368,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 1.0233965555826823,
+      "learning_rate": 0.0001,
+      "loss": 3.1617,
+      "loss/crossentropy": 2.3866767287254333,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25886829346418383,
+      "loss/reg": 0.0,
+      "step": 17180
+    },
+    {
+      "epoch": 0.1130921052631579,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.08580322265625,
+      "learning_rate": 0.0001,
+      "loss": 3.1277,
+      "loss/crossentropy": 2.095993900299072,
+      "loss/hidden": 2.6625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1956949472427368,
+      "loss/reg": 0.0,
+      "step": 17190
+    },
+    {
+      "epoch": 0.11315789473684211,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.09143778483072916,
+      "learning_rate": 0.0001,
+      "loss": 3.0835,
+      "loss/crossentropy": 2.3784751892089844,
+      "loss/hidden": 2.809375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26899502128362657,
+      "loss/reg": 0.0,
+      "step": 17200
+    },
+    {
+      "epoch": 0.11322368421052631,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.027274576822916667,
+      "learning_rate": 0.0001,
+      "loss": 3.1617,
+      "loss/crossentropy": 2.505306875705719,
+      "loss/hidden": 2.953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2929636001586914,
+      "loss/reg": 0.0,
+      "step": 17210
+    },
+    {
+      "epoch": 0.11328947368421052,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.06923421223958333,
+      "learning_rate": 0.0001,
+      "loss": 3.12,
+      "loss/crossentropy": 2.1889270186424254,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2669166073203087,
+      "loss/reg": 0.0,
+      "step": 17220
+    },
+    {
+      "epoch": 0.11335526315789474,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.17213312784830728,
+      "learning_rate": 0.0001,
+      "loss": 3.0681,
+      "loss/crossentropy": 2.2528875708580016,
+      "loss/hidden": 2.7375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23936019986867904,
+      "loss/reg": 0.0,
+      "step": 17230
+    },
+    {
+      "epoch": 0.11342105263157895,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.12957763671875,
+      "learning_rate": 0.0001,
+      "loss": 3.119,
+      "loss/crossentropy": 2.2457367897033693,
+      "loss/hidden": 2.890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23305099308490754,
+      "loss/reg": 0.0,
+      "step": 17240
+    },
+    {
+      "epoch": 0.11348684210526316,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.07983779907226562,
+      "learning_rate": 0.0001,
+      "loss": 3.0851,
+      "loss/crossentropy": 2.4546321392059327,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24581009149551392,
+      "loss/reg": 0.0,
+      "step": 17250
+    },
+    {
+      "epoch": 0.11355263157894736,
+      "grad_norm": 4.1875,
+      "grad_norm_var": 0.2813791910807292,
+      "learning_rate": 0.0001,
+      "loss": 3.171,
+      "loss/crossentropy": 1.9349523544311524,
+      "loss/hidden": 3.0453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31686680018901825,
+      "loss/reg": 0.0,
+      "step": 17260
+    },
+    {
+      "epoch": 0.11361842105263158,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 1.4602701822916666,
+      "learning_rate": 0.0001,
+      "loss": 3.1656,
+      "loss/crossentropy": 2.2751689314842225,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2878335312008858,
+      "loss/reg": 0.0,
+      "step": 17270
+    },
+    {
+      "epoch": 0.11368421052631579,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.5687662760416666,
+      "learning_rate": 0.0001,
+      "loss": 3.0772,
+      "loss/crossentropy": 2.4036699175834655,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24233294725418092,
+      "loss/reg": 0.0,
+      "step": 17280
+    },
+    {
+      "epoch": 0.11375,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.06691792805989584,
+      "learning_rate": 0.0001,
+      "loss": 3.1161,
+      "loss/crossentropy": 2.4056329488754273,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25776179879903793,
+      "loss/reg": 0.0,
+      "step": 17290
+    },
+    {
+      "epoch": 0.11381578947368422,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.12681884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.2097,
+      "loss/crossentropy": 2.4936662912368774,
+      "loss/hidden": 2.9015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27261604368686676,
+      "loss/reg": 0.0,
+      "step": 17300
+    },
+    {
+      "epoch": 0.11388157894736842,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.08603108723958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1596,
+      "loss/crossentropy": 2.2334436774253845,
+      "loss/hidden": 2.978125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27515108734369276,
+      "loss/reg": 0.0,
+      "step": 17310
+    },
+    {
+      "epoch": 0.11394736842105263,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.06290690104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1322,
+      "loss/crossentropy": 2.3009248971939087,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2514318063855171,
+      "loss/reg": 0.0,
+      "step": 17320
+    },
+    {
+      "epoch": 0.11401315789473684,
+      "grad_norm": 2.953125,
+      "grad_norm_var": 0.19062093098958333,
+      "learning_rate": 0.0001,
+      "loss": 3.2222,
+      "loss/crossentropy": 2.2115599513053894,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24124337881803512,
+      "loss/reg": 0.0,
+      "step": 17330
+    },
+    {
+      "epoch": 0.11407894736842106,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.4003245035807292,
+      "learning_rate": 0.0001,
+      "loss": 3.1404,
+      "loss/crossentropy": 2.2533546447753907,
+      "loss/hidden": 3.0046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25740948766469957,
+      "loss/reg": 0.0,
+      "step": 17340
+    },
+    {
+      "epoch": 0.11414473684210526,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.06924540201822917,
+      "learning_rate": 0.0001,
+      "loss": 3.1509,
+      "loss/crossentropy": 2.4188032507896424,
+      "loss/hidden": 2.9,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27805479913949965,
+      "loss/reg": 0.0,
+      "step": 17350
+    },
+    {
+      "epoch": 0.11421052631578947,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.6562489827473958,
+      "learning_rate": 0.0001,
+      "loss": 3.2795,
+      "loss/crossentropy": 2.1555517435073854,
+      "loss/hidden": 2.925,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28488118648529054,
+      "loss/reg": 0.0,
+      "step": 17360
+    },
+    {
+      "epoch": 0.11427631578947368,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.62613525390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1637,
+      "loss/crossentropy": 2.351659083366394,
+      "loss/hidden": 2.959375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2672264903783798,
+      "loss/reg": 0.0,
+      "step": 17370
+    },
+    {
+      "epoch": 0.1143421052631579,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.08678385416666666,
+      "learning_rate": 0.0001,
+      "loss": 3.1242,
+      "loss/crossentropy": 2.2498091578483583,
+      "loss/hidden": 2.8359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24537185132503508,
+      "loss/reg": 0.0,
+      "step": 17380
+    },
+    {
+      "epoch": 0.11440789473684211,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.17229817708333334,
+      "learning_rate": 0.0001,
+      "loss": 3.1776,
+      "loss/crossentropy": 2.401424062252045,
+      "loss/hidden": 3.0328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3059739723801613,
+      "loss/reg": 0.0,
+      "step": 17390
+    },
+    {
+      "epoch": 0.11447368421052631,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.1625885009765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1477,
+      "loss/crossentropy": 2.292530918121338,
+      "loss/hidden": 3.09375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2784269869327545,
+      "loss/reg": 0.0,
+      "step": 17400
+    },
+    {
+      "epoch": 0.11453947368421052,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.09892171223958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1288,
+      "loss/crossentropy": 2.464983069896698,
+      "loss/hidden": 2.9078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2796273499727249,
+      "loss/reg": 0.0,
+      "step": 17410
+    },
+    {
+      "epoch": 0.11460526315789474,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.09850972493489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1289,
+      "loss/crossentropy": 2.1951366662979126,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21696581244468688,
+      "loss/reg": 0.0,
+      "step": 17420
+    },
+    {
+      "epoch": 0.11467105263157895,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.06843159993489584,
+      "learning_rate": 0.0001,
+      "loss": 3.1597,
+      "loss/crossentropy": 2.3925849914550783,
+      "loss/hidden": 2.9625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2907762542366982,
+      "loss/reg": 0.0,
+      "step": 17430
+    },
+    {
+      "epoch": 0.11473684210526315,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.21046549479166668,
+      "learning_rate": 0.0001,
+      "loss": 3.1634,
+      "loss/crossentropy": 2.464526927471161,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2937367483973503,
+      "loss/reg": 0.0,
+      "step": 17440
+    },
+    {
+      "epoch": 0.11480263157894736,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.047526041666666664,
+      "learning_rate": 0.0001,
+      "loss": 3.1144,
+      "loss/crossentropy": 2.072342965006828,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2236027292907238,
+      "loss/reg": 0.0,
+      "step": 17450
+    },
+    {
+      "epoch": 0.11486842105263158,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.05799051920572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0873,
+      "loss/crossentropy": 2.240265655517578,
+      "loss/hidden": 2.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23422597646713256,
+      "loss/reg": 0.0,
+      "step": 17460
+    },
+    {
+      "epoch": 0.11493421052631579,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.10058186848958334,
+      "learning_rate": 0.0001,
+      "loss": 3.127,
+      "loss/crossentropy": 2.3168819308280946,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2430237874388695,
+      "loss/reg": 0.0,
+      "step": 17470
+    },
+    {
+      "epoch": 0.115,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.11015523274739583,
+      "learning_rate": 0.0001,
+      "loss": 3.0618,
+      "loss/crossentropy": 2.2366016268730164,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23987135738134385,
+      "loss/reg": 0.0,
+      "step": 17480
+    },
+    {
+      "epoch": 0.1150657894736842,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.09355061848958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1875,
+      "loss/crossentropy": 2.1879064559936525,
+      "loss/hidden": 3.1921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24197361022233962,
+      "loss/reg": 0.0,
+      "step": 17490
+    },
+    {
+      "epoch": 0.11513157894736842,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 7.735667928059896,
+      "learning_rate": 0.0001,
+      "loss": 3.2427,
+      "loss/crossentropy": 2.149078643321991,
+      "loss/hidden": 3.009375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24705443456768988,
+      "loss/reg": 0.0,
+      "step": 17500
+    },
+    {
+      "epoch": 0.11519736842105263,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 7.899019368489584,
+      "learning_rate": 0.0001,
+      "loss": 3.1837,
+      "loss/crossentropy": 2.3580456256866453,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23221501410007478,
+      "loss/reg": 0.0,
+      "step": 17510
+    },
+    {
+      "epoch": 0.11526315789473685,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.2678456624348958,
+      "learning_rate": 0.0001,
+      "loss": 3.079,
+      "loss/crossentropy": 2.414518404006958,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2647902578115463,
+      "loss/reg": 0.0,
+      "step": 17520
+    },
+    {
+      "epoch": 0.11532894736842106,
+      "grad_norm": 3.125,
+      "grad_norm_var": 0.14198811848958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1613,
+      "loss/crossentropy": 2.416664445400238,
+      "loss/hidden": 2.7625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22963229715824127,
+      "loss/reg": 0.0,
+      "step": 17530
+    },
+    {
+      "epoch": 0.11539473684210526,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.7926177978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1532,
+      "loss/crossentropy": 2.1853476405143737,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2286560907959938,
+      "loss/reg": 0.0,
+      "step": 17540
+    },
+    {
+      "epoch": 0.11546052631578947,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.8031534830729167,
+      "learning_rate": 0.0001,
+      "loss": 3.1557,
+      "loss/crossentropy": 2.6209848165512084,
+      "loss/hidden": 2.7359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25361735969781873,
+      "loss/reg": 0.0,
+      "step": 17550
+    },
+    {
+      "epoch": 0.11552631578947369,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.09949544270833334,
+      "learning_rate": 0.0001,
+      "loss": 3.2229,
+      "loss/crossentropy": 2.3588363409042357,
+      "loss/hidden": 2.978125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30205955654382705,
+      "loss/reg": 0.0,
+      "step": 17560
+    },
+    {
+      "epoch": 0.1155921052631579,
+      "grad_norm": 1.9765625,
+      "grad_norm_var": 0.07576471964518229,
+      "learning_rate": 0.0001,
+      "loss": 3.174,
+      "loss/crossentropy": 2.1386860758066177,
+      "loss/hidden": 2.871875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24600727967917918,
+      "loss/reg": 0.0,
+      "step": 17570
+    },
+    {
+      "epoch": 0.1156578947368421,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.07134577433268229,
+      "learning_rate": 0.0001,
+      "loss": 3.1143,
+      "loss/crossentropy": 2.5142947912216185,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2483747810125351,
+      "loss/reg": 0.0,
+      "step": 17580
+    },
+    {
+      "epoch": 0.11572368421052631,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.7249959309895834,
+      "learning_rate": 0.0001,
+      "loss": 3.1598,
+      "loss/crossentropy": 2.339717888832092,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2511411294341087,
+      "loss/reg": 0.0,
+      "step": 17590
+    },
+    {
+      "epoch": 0.11578947368421053,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.9438435872395833,
+      "learning_rate": 0.0001,
+      "loss": 3.0902,
+      "loss/crossentropy": 2.3273319840431212,
+      "loss/hidden": 2.9015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.321346378326416,
+      "loss/reg": 0.0,
+      "step": 17600
+    },
+    {
+      "epoch": 0.11585526315789474,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.93150634765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1392,
+      "loss/crossentropy": 2.0881393194198608,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2542056769132614,
+      "loss/reg": 0.0,
+      "step": 17610
+    },
+    {
+      "epoch": 0.11592105263157895,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.05450846354166667,
+      "learning_rate": 0.0001,
+      "loss": 3.153,
+      "loss/crossentropy": 2.200914776325226,
+      "loss/hidden": 2.9015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2596256859600544,
+      "loss/reg": 0.0,
+      "step": 17620
+    },
+    {
+      "epoch": 0.11598684210526315,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.06134440104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1138,
+      "loss/crossentropy": 2.391866648197174,
+      "loss/hidden": 2.7578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24801254719495774,
+      "loss/reg": 0.0,
+      "step": 17630
+    },
+    {
+      "epoch": 0.11605263157894737,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.19595947265625,
+      "learning_rate": 0.0001,
+      "loss": 3.1197,
+      "loss/crossentropy": 2.3731101989746093,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2734734550118446,
+      "loss/reg": 0.0,
+      "step": 17640
+    },
+    {
+      "epoch": 0.11611842105263158,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.06402587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.186,
+      "loss/crossentropy": 2.510706162452698,
+      "loss/hidden": 3.1234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3243511900305748,
+      "loss/reg": 0.0,
+      "step": 17650
+    },
+    {
+      "epoch": 0.11618421052631579,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.08347066243489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1433,
+      "loss/crossentropy": 2.3694114327430724,
+      "loss/hidden": 2.8859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23816563338041305,
+      "loss/reg": 0.0,
+      "step": 17660
+    },
+    {
+      "epoch": 0.11625,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.07176106770833333,
+      "learning_rate": 0.0001,
+      "loss": 3.117,
+      "loss/crossentropy": 2.3822181940078737,
+      "loss/hidden": 2.90625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2559090554714203,
+      "loss/reg": 0.0,
+      "step": 17670
+    },
+    {
+      "epoch": 0.1163157894736842,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.06597391764322917,
+      "learning_rate": 0.0001,
+      "loss": 3.167,
+      "loss/crossentropy": 2.236483609676361,
+      "loss/hidden": 2.753125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24177847057580948,
+      "loss/reg": 0.0,
+      "step": 17680
+    },
+    {
+      "epoch": 0.11638157894736842,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.05454813639322917,
+      "learning_rate": 0.0001,
+      "loss": 3.1207,
+      "loss/crossentropy": 2.386284852027893,
+      "loss/hidden": 2.725,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23972099274396896,
+      "loss/reg": 0.0,
+      "step": 17690
+    },
+    {
+      "epoch": 0.11644736842105263,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.029325358072916665,
+      "learning_rate": 0.0001,
+      "loss": 3.1551,
+      "loss/crossentropy": 2.3102688074111937,
+      "loss/hidden": 2.9453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25639690458774567,
+      "loss/reg": 0.0,
+      "step": 17700
+    },
+    {
+      "epoch": 0.11651315789473685,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.1159576416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1227,
+      "loss/crossentropy": 2.164554476737976,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2502566508948803,
+      "loss/reg": 0.0,
+      "step": 17710
+    },
+    {
+      "epoch": 0.11657894736842105,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.16982421875,
+      "learning_rate": 0.0001,
+      "loss": 3.2736,
+      "loss/crossentropy": 2.166772598028183,
+      "loss/hidden": 3.0484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24817814379930497,
+      "loss/reg": 0.0,
+      "step": 17720
+    },
+    {
+      "epoch": 0.11664473684210526,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.13912353515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1985,
+      "loss/crossentropy": 2.1537662744522095,
+      "loss/hidden": 2.821875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21837295591831207,
+      "loss/reg": 0.0,
+      "step": 17730
+    },
+    {
+      "epoch": 0.11671052631578947,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.09753392537434896,
+      "learning_rate": 0.0001,
+      "loss": 3.1066,
+      "loss/crossentropy": 2.106878674030304,
+      "loss/hidden": 2.840625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2489032343029976,
+      "loss/reg": 0.0,
+      "step": 17740
+    },
+    {
+      "epoch": 0.11677631578947369,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.08820699055989584,
+      "learning_rate": 0.0001,
+      "loss": 3.1647,
+      "loss/crossentropy": 2.186918389797211,
+      "loss/hidden": 3.1171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26816043853759763,
+      "loss/reg": 0.0,
+      "step": 17750
+    },
+    {
+      "epoch": 0.1168421052631579,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.14002176920572917,
+      "learning_rate": 0.0001,
+      "loss": 3.2041,
+      "loss/crossentropy": 2.130704140663147,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26593275666236876,
+      "loss/reg": 0.0,
+      "step": 17760
+    },
+    {
+      "epoch": 0.1169078947368421,
+      "grad_norm": 3.6875,
+      "grad_norm_var": 0.16592508951822918,
+      "learning_rate": 0.0001,
+      "loss": 3.0748,
+      "loss/crossentropy": 2.4294887661933897,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26008370518684387,
+      "loss/reg": 0.0,
+      "step": 17770
+    },
+    {
+      "epoch": 0.11697368421052631,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.4354156494140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2129,
+      "loss/crossentropy": 2.0633517861366273,
+      "loss/hidden": 2.9359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2134759709239006,
+      "loss/reg": 0.0,
+      "step": 17780
+    },
+    {
+      "epoch": 0.11703947368421053,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.114306640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1969,
+      "loss/crossentropy": 2.470662558078766,
+      "loss/hidden": 2.9390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32178338468074796,
+      "loss/reg": 0.0,
+      "step": 17790
+    },
+    {
+      "epoch": 0.11710526315789474,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.1042388916015625,
+      "learning_rate": 0.0001,
+      "loss": 3.132,
+      "loss/crossentropy": 2.180730104446411,
+      "loss/hidden": 2.8953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2408156231045723,
+      "loss/reg": 0.0,
+      "step": 17800
+    },
+    {
+      "epoch": 0.11717105263157895,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.08092041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1107,
+      "loss/crossentropy": 2.3672548174858092,
+      "loss/hidden": 2.8234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2295123293995857,
+      "loss/reg": 0.0,
+      "step": 17810
+    },
+    {
+      "epoch": 0.11723684210526315,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.029108683268229168,
+      "learning_rate": 0.0001,
+      "loss": 3.1063,
+      "loss/crossentropy": 2.486844336986542,
+      "loss/hidden": 2.921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2817386701703072,
+      "loss/reg": 0.0,
+      "step": 17820
+    },
+    {
+      "epoch": 0.11730263157894737,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 3.6479156901264755e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2511,
+      "loss/crossentropy": 2.3458006739616395,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25344357788562777,
+      "loss/reg": 0.0,
+      "step": 17830
+    },
+    {
+      "epoch": 0.11736842105263158,
+      "grad_norm": 2.75,
+      "grad_norm_var": 1.1187174479166666,
+      "learning_rate": 0.0001,
+      "loss": 3.1952,
+      "loss/crossentropy": 2.3346620917320253,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24867950975894929,
+      "loss/reg": 0.0,
+      "step": 17840
+    },
+    {
+      "epoch": 0.1174342105263158,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.05538736979166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1781,
+      "loss/crossentropy": 2.1594719171524046,
+      "loss/hidden": 2.9390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32041922956705093,
+      "loss/reg": 0.0,
+      "step": 17850
+    },
+    {
+      "epoch": 0.1175,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.13107096354166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1675,
+      "loss/crossentropy": 2.335154187679291,
+      "loss/hidden": 2.653125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23140522688627244,
+      "loss/reg": 0.0,
+      "step": 17860
+    },
+    {
+      "epoch": 0.1175657894736842,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.061777496337890626,
+      "learning_rate": 0.0001,
+      "loss": 3.1431,
+      "loss/crossentropy": 2.448777401447296,
+      "loss/hidden": 2.78125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24006084948778153,
+      "loss/reg": 0.0,
+      "step": 17870
+    },
+    {
+      "epoch": 0.11763157894736842,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.14251302083333334,
+      "learning_rate": 0.0001,
+      "loss": 3.2176,
+      "loss/crossentropy": 2.2756917595863344,
+      "loss/hidden": 3.0359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30769334733486176,
+      "loss/reg": 0.0,
+      "step": 17880
+    },
+    {
+      "epoch": 0.11769736842105263,
+      "grad_norm": 1.921875,
+      "grad_norm_var": 0.14058837890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1661,
+      "loss/crossentropy": 2.3477345585823057,
+      "loss/hidden": 2.740625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2401590347290039,
+      "loss/reg": 0.0,
+      "step": 17890
+    },
+    {
+      "epoch": 0.11776315789473685,
+      "grad_norm": 3.34375,
+      "grad_norm_var": 0.14404271443684896,
+      "learning_rate": 0.0001,
+      "loss": 3.1339,
+      "loss/crossentropy": 2.2531643748283385,
+      "loss/hidden": 2.9875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27347354739904406,
+      "loss/reg": 0.0,
+      "step": 17900
+    },
+    {
+      "epoch": 0.11782894736842105,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.10000991821289062,
+      "learning_rate": 0.0001,
+      "loss": 3.1182,
+      "loss/crossentropy": 2.2883235216140747,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2531774565577507,
+      "loss/reg": 0.0,
+      "step": 17910
+    },
+    {
+      "epoch": 0.11789473684210526,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.0377349853515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1566,
+      "loss/crossentropy": 2.2789531648159027,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.235878374427557,
+      "loss/reg": 0.0,
+      "step": 17920
+    },
+    {
+      "epoch": 0.11796052631578947,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.06096598307291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1068,
+      "loss/crossentropy": 2.3758100509643554,
+      "loss/hidden": 2.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22890822887420653,
+      "loss/reg": 0.0,
+      "step": 17930
+    },
+    {
+      "epoch": 0.11802631578947369,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.07280171712239583,
+      "learning_rate": 0.0001,
+      "loss": 3.1773,
+      "loss/crossentropy": 2.2372307777404785,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24057336896657944,
+      "loss/reg": 0.0,
+      "step": 17940
+    },
+    {
+      "epoch": 0.1180921052631579,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.12463785807291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1969,
+      "loss/crossentropy": 2.1488450884819033,
+      "loss/hidden": 2.9078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25413042977452277,
+      "loss/reg": 0.0,
+      "step": 17950
+    },
+    {
+      "epoch": 0.1181578947368421,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.11864827473958334,
+      "learning_rate": 0.0001,
+      "loss": 3.2164,
+      "loss/crossentropy": 2.5382091283798216,
+      "loss/hidden": 3.13125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26207938939332964,
+      "loss/reg": 0.0,
+      "step": 17960
+    },
+    {
+      "epoch": 0.11822368421052631,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.06974283854166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1713,
+      "loss/crossentropy": 2.3174261093139648,
+      "loss/hidden": 2.9828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.279338338971138,
+      "loss/reg": 0.0,
+      "step": 17970
+    },
+    {
+      "epoch": 0.11828947368421053,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.04395243326822917,
+      "learning_rate": 0.0001,
+      "loss": 3.1261,
+      "loss/crossentropy": 2.182445216178894,
+      "loss/hidden": 2.8578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2768437474966049,
+      "loss/reg": 0.0,
+      "step": 17980
+    },
+    {
+      "epoch": 0.11835526315789474,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.14431050618489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1287,
+      "loss/crossentropy": 2.202815556526184,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21700907945632936,
+      "loss/reg": 0.0,
+      "step": 17990
+    },
+    {
+      "epoch": 0.11842105263157894,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 2.2837799072265623,
+      "learning_rate": 0.0001,
+      "loss": 3.1737,
+      "loss/crossentropy": 2.0051895678043365,
+      "loss/hidden": 3.2046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.35186032503843306,
+      "loss/reg": 0.0,
+      "step": 18000
+    },
+    {
+      "epoch": 0.11848684210526315,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 2.469374338785807,
+      "learning_rate": 0.0001,
+      "loss": 3.0848,
+      "loss/crossentropy": 1.872368621826172,
+      "loss/hidden": 2.9765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2913415163755417,
+      "loss/reg": 0.0,
+      "step": 18010
+    },
+    {
+      "epoch": 0.11855263157894737,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.1679278055826823,
+      "learning_rate": 0.0001,
+      "loss": 3.2063,
+      "loss/crossentropy": 2.367658519744873,
+      "loss/hidden": 2.95,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27947854697704316,
+      "loss/reg": 0.0,
+      "step": 18020
+    },
+    {
+      "epoch": 0.11861842105263158,
+      "grad_norm": 3.40625,
+      "grad_norm_var": 0.15479227701822917,
+      "learning_rate": 0.0001,
+      "loss": 3.1836,
+      "loss/crossentropy": 2.087648892402649,
+      "loss/hidden": 2.8875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2204158440232277,
+      "loss/reg": 0.0,
+      "step": 18030
+    },
+    {
+      "epoch": 0.1186842105263158,
+      "grad_norm": 3.171875,
+      "grad_norm_var": 0.18935546875,
+      "learning_rate": 0.0001,
+      "loss": 3.205,
+      "loss/crossentropy": 2.1874751687049865,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2435563921928406,
+      "loss/reg": 0.0,
+      "step": 18040
+    },
+    {
+      "epoch": 0.11875,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.38063151041666665,
+      "learning_rate": 0.0001,
+      "loss": 3.1996,
+      "loss/crossentropy": 2.1633397549390794,
+      "loss/hidden": 2.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21968650221824645,
+      "loss/reg": 0.0,
+      "step": 18050
+    },
+    {
+      "epoch": 0.11881578947368421,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.5637278238932292,
+      "learning_rate": 0.0001,
+      "loss": 3.1736,
+      "loss/crossentropy": 2.281619644165039,
+      "loss/hidden": 2.93125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.295219412446022,
+      "loss/reg": 0.0,
+      "step": 18060
+    },
+    {
+      "epoch": 0.11888157894736842,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 1.9192342122395833,
+      "learning_rate": 0.0001,
+      "loss": 3.2705,
+      "loss/crossentropy": 2.21823273897171,
+      "loss/hidden": 2.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23566214740276337,
+      "loss/reg": 0.0,
+      "step": 18070
+    },
+    {
+      "epoch": 0.11894736842105263,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 1.6334869384765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1873,
+      "loss/crossentropy": 2.2668980836868284,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2256957158446312,
+      "loss/reg": 0.0,
+      "step": 18080
+    },
+    {
+      "epoch": 0.11901315789473685,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.0713287353515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0942,
+      "loss/crossentropy": 2.218550479412079,
+      "loss/hidden": 3.0296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3167494982481003,
+      "loss/reg": 0.0,
+      "step": 18090
+    },
+    {
+      "epoch": 0.11907894736842105,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.05681050618489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1816,
+      "loss/crossentropy": 2.464331579208374,
+      "loss/hidden": 2.984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25844376236200334,
+      "loss/reg": 0.0,
+      "step": 18100
+    },
+    {
+      "epoch": 0.11914473684210526,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.07647196451822917,
+      "learning_rate": 0.0001,
+      "loss": 3.1727,
+      "loss/crossentropy": 2.2838521599769592,
+      "loss/hidden": 2.94375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2587725341320038,
+      "loss/reg": 0.0,
+      "step": 18110
+    },
+    {
+      "epoch": 0.11921052631578948,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.13923238118489584,
+      "learning_rate": 0.0001,
+      "loss": 3.2584,
+      "loss/crossentropy": 2.3038102626800536,
+      "loss/hidden": 3.0203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3060797408223152,
+      "loss/reg": 0.0,
+      "step": 18120
+    },
+    {
+      "epoch": 0.11927631578947369,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.11160481770833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0877,
+      "loss/crossentropy": 2.1922419667243958,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2569601759314537,
+      "loss/reg": 0.0,
+      "step": 18130
+    },
+    {
+      "epoch": 0.11934210526315789,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.18005269368489582,
+      "learning_rate": 0.0001,
+      "loss": 3.2654,
+      "loss/crossentropy": 2.336120533943176,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3214158996939659,
+      "loss/reg": 0.0,
+      "step": 18140
+    },
+    {
+      "epoch": 0.1194078947368421,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.211083984375,
+      "learning_rate": 0.0001,
+      "loss": 3.1801,
+      "loss/crossentropy": 2.6258700489997864,
+      "loss/hidden": 2.9328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23279473930597305,
+      "loss/reg": 0.0,
+      "step": 18150
+    },
+    {
+      "epoch": 0.11947368421052632,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.097412109375,
+      "learning_rate": 0.0001,
+      "loss": 3.0988,
+      "loss/crossentropy": 2.378826451301575,
+      "loss/hidden": 2.709375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24316011592745781,
+      "loss/reg": 0.0,
+      "step": 18160
+    },
+    {
+      "epoch": 0.11953947368421053,
+      "grad_norm": 2.953125,
+      "grad_norm_var": 0.25974934895833335,
+      "learning_rate": 0.0001,
+      "loss": 3.1615,
+      "loss/crossentropy": 2.3753814458847047,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2662284314632416,
+      "loss/reg": 0.0,
+      "step": 18170
+    },
+    {
+      "epoch": 0.11960526315789474,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.2468658447265625,
+      "learning_rate": 0.0001,
+      "loss": 3.2061,
+      "loss/crossentropy": 2.2813811898231506,
+      "loss/hidden": 3.0953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3734820380806923,
+      "loss/reg": 0.0,
+      "step": 18180
+    },
+    {
+      "epoch": 0.11967105263157894,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.12701416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.16,
+      "loss/crossentropy": 2.2644060015678407,
+      "loss/hidden": 2.9109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22710230052471161,
+      "loss/reg": 0.0,
+      "step": 18190
+    },
+    {
+      "epoch": 0.11973684210526316,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.059342447916666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0907,
+      "loss/crossentropy": 2.2262014031410216,
+      "loss/hidden": 2.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20997287034988404,
+      "loss/reg": 0.0,
+      "step": 18200
+    },
+    {
+      "epoch": 0.11980263157894737,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.04023335774739583,
+      "learning_rate": 0.0001,
+      "loss": 3.2103,
+      "loss/crossentropy": 2.261046063899994,
+      "loss/hidden": 2.996875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.262168163061142,
+      "loss/reg": 0.0,
+      "step": 18210
+    },
+    {
+      "epoch": 0.11986842105263158,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.11513570149739584,
+      "learning_rate": 0.0001,
+      "loss": 3.194,
+      "loss/crossentropy": 2.39845809340477,
+      "loss/hidden": 2.9171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2080080732703209,
+      "loss/reg": 0.0,
+      "step": 18220
+    },
+    {
+      "epoch": 0.1199342105263158,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.12336324055989584,
+      "learning_rate": 0.0001,
+      "loss": 3.1321,
+      "loss/crossentropy": 2.221972668170929,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2630168259143829,
+      "loss/reg": 0.0,
+      "step": 18230
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.0994049072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.192,
+      "loss/crossentropy": 1.910440945625305,
+      "loss/hidden": 2.9625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22247039675712585,
+      "loss/reg": 0.0,
+      "step": 18240
+    },
+    {
+      "epoch": 0.12006578947368421,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.38481343587239586,
+      "learning_rate": 0.0001,
+      "loss": 3.211,
+      "loss/crossentropy": 2.3357484221458433,
+      "loss/hidden": 3.1015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28341811895370483,
+      "loss/reg": 0.0,
+      "step": 18250
+    },
+    {
+      "epoch": 0.12013157894736842,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.1193023681640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1537,
+      "loss/crossentropy": 2.277235043048859,
+      "loss/hidden": 3.003125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2401238664984703,
+      "loss/reg": 0.0,
+      "step": 18260
+    },
+    {
+      "epoch": 0.12019736842105264,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.0804595947265625,
+      "learning_rate": 0.0001,
+      "loss": 3.1693,
+      "loss/crossentropy": 2.4965412855148315,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23239507675170898,
+      "loss/reg": 0.0,
+      "step": 18270
+    },
+    {
+      "epoch": 0.12026315789473684,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.1340240478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0471,
+      "loss/crossentropy": 2.187135934829712,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22307218313217164,
+      "loss/reg": 0.0,
+      "step": 18280
+    },
+    {
+      "epoch": 0.12032894736842105,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.08740946451822916,
+      "learning_rate": 0.0001,
+      "loss": 3.0703,
+      "loss/crossentropy": 2.144671416282654,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23641343265771866,
+      "loss/reg": 0.0,
+      "step": 18290
+    },
+    {
+      "epoch": 0.12039473684210526,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.0469390869140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0766,
+      "loss/crossentropy": 2.3019802451133726,
+      "loss/hidden": 2.909375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27486053854227066,
+      "loss/reg": 0.0,
+      "step": 18300
+    },
+    {
+      "epoch": 0.12046052631578948,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.056086222330729164,
+      "learning_rate": 0.0001,
+      "loss": 3.1479,
+      "loss/crossentropy": 2.350484275817871,
+      "loss/hidden": 3.0453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2915784493088722,
+      "loss/reg": 0.0,
+      "step": 18310
+    },
+    {
+      "epoch": 0.12052631578947369,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.0907867431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.085,
+      "loss/crossentropy": 2.519459903240204,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2799163952469826,
+      "loss/reg": 0.0,
+      "step": 18320
+    },
+    {
+      "epoch": 0.12059210526315789,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.07711588541666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1352,
+      "loss/crossentropy": 2.196255683898926,
+      "loss/hidden": 3.0125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.256213016808033,
+      "loss/reg": 0.0,
+      "step": 18330
+    },
+    {
+      "epoch": 0.1206578947368421,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.111083984375,
+      "learning_rate": 0.0001,
+      "loss": 3.1252,
+      "loss/crossentropy": 2.421483266353607,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3222003743052483,
+      "loss/reg": 0.0,
+      "step": 18340
+    },
+    {
+      "epoch": 0.12072368421052632,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.07574462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1338,
+      "loss/crossentropy": 2.248370945453644,
+      "loss/hidden": 2.86875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24929146319627762,
+      "loss/reg": 0.0,
+      "step": 18350
+    },
+    {
+      "epoch": 0.12078947368421053,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.05138346354166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1477,
+      "loss/crossentropy": 2.3680251955986025,
+      "loss/hidden": 3.021875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2555671989917755,
+      "loss/reg": 0.0,
+      "step": 18360
+    },
+    {
+      "epoch": 0.12085526315789474,
+      "grad_norm": 3.296875,
+      "grad_norm_var": 0.06642964680989584,
+      "learning_rate": 0.0001,
+      "loss": 3.1165,
+      "loss/crossentropy": 2.199974000453949,
+      "loss/hidden": 2.86875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24803201854228973,
+      "loss/reg": 0.0,
+      "step": 18370
+    },
+    {
+      "epoch": 0.12092105263157894,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.11695556640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1367,
+      "loss/crossentropy": 2.5192595601081846,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24734148681163787,
+      "loss/reg": 0.0,
+      "step": 18380
+    },
+    {
+      "epoch": 0.12098684210526316,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.05526936848958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1724,
+      "loss/crossentropy": 2.450891613960266,
+      "loss/hidden": 2.9140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2809787794947624,
+      "loss/reg": 0.0,
+      "step": 18390
+    },
+    {
+      "epoch": 0.12105263157894737,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.0400787353515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1368,
+      "loss/crossentropy": 2.142946255207062,
+      "loss/hidden": 2.9015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25829449892044065,
+      "loss/reg": 0.0,
+      "step": 18400
+    },
+    {
+      "epoch": 0.12111842105263158,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.12823893229166666,
+      "learning_rate": 0.0001,
+      "loss": 3.1313,
+      "loss/crossentropy": 2.4548865795135497,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2860646352171898,
+      "loss/reg": 0.0,
+      "step": 18410
+    },
+    {
+      "epoch": 0.12118421052631578,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.22473042805989582,
+      "learning_rate": 0.0001,
+      "loss": 3.0336,
+      "loss/crossentropy": 2.4296185970306396,
+      "loss/hidden": 2.8359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24130426943302155,
+      "loss/reg": 0.0,
+      "step": 18420
+    },
+    {
+      "epoch": 0.12125,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.24602864583333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0905,
+      "loss/crossentropy": 2.2097928047180178,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22747812122106553,
+      "loss/reg": 0.0,
+      "step": 18430
+    },
+    {
+      "epoch": 0.12131578947368421,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.03878580729166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1482,
+      "loss/crossentropy": 2.266276228427887,
+      "loss/hidden": 3.1015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3076672673225403,
+      "loss/reg": 0.0,
+      "step": 18440
+    },
+    {
+      "epoch": 0.12138157894736842,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.19593098958333333,
+      "learning_rate": 0.0001,
+      "loss": 3.1753,
+      "loss/crossentropy": 2.3656784892082214,
+      "loss/hidden": 2.634375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21082431972026824,
+      "loss/reg": 0.0,
+      "step": 18450
+    },
+    {
+      "epoch": 0.12144736842105264,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.13351236979166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1809,
+      "loss/crossentropy": 2.363306760787964,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.259796117246151,
+      "loss/reg": 0.0,
+      "step": 18460
+    },
+    {
+      "epoch": 0.12151315789473684,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.10623372395833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0904,
+      "loss/crossentropy": 2.2991411328315734,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22767046988010406,
+      "loss/reg": 0.0,
+      "step": 18470
+    },
+    {
+      "epoch": 0.12157894736842105,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.051595052083333336,
+      "learning_rate": 0.0001,
+      "loss": 3.0997,
+      "loss/crossentropy": 2.479309868812561,
+      "loss/hidden": 2.703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2216094933450222,
+      "loss/reg": 0.0,
+      "step": 18480
+    },
+    {
+      "epoch": 0.12164473684210526,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.029279581705729165,
+      "learning_rate": 0.0001,
+      "loss": 3.0982,
+      "loss/crossentropy": 2.224264907836914,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2954641401767731,
+      "loss/reg": 0.0,
+      "step": 18490
+    },
+    {
+      "epoch": 0.12171052631578948,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.0418365478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1278,
+      "loss/crossentropy": 2.1996599078178405,
+      "loss/hidden": 2.740625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24900650084018708,
+      "loss/reg": 0.0,
+      "step": 18500
+    },
+    {
+      "epoch": 0.12177631578947369,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 6.255301920572917,
+      "learning_rate": 0.0001,
+      "loss": 3.1924,
+      "loss/crossentropy": 2.388622558116913,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2473811611533165,
+      "loss/reg": 0.0,
+      "step": 18510
+    },
+    {
+      "epoch": 0.12184210526315789,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.07906901041666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1227,
+      "loss/crossentropy": 2.325872230529785,
+      "loss/hidden": 2.9140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2705776423215866,
+      "loss/reg": 0.0,
+      "step": 18520
+    },
+    {
+      "epoch": 0.1219078947368421,
+      "grad_norm": 3.96875,
+      "grad_norm_var": 0.2271197001139323,
+      "learning_rate": 0.0001,
+      "loss": 3.104,
+      "loss/crossentropy": 2.4012768149375914,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24956294745206833,
+      "loss/reg": 0.0,
+      "step": 18530
+    },
+    {
+      "epoch": 0.12197368421052632,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.2910316467285156,
+      "learning_rate": 0.0001,
+      "loss": 3.1206,
+      "loss/crossentropy": 2.605260455608368,
+      "loss/hidden": 3.33125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2743007704615593,
+      "loss/reg": 0.0,
+      "step": 18540
+    },
+    {
+      "epoch": 0.12203947368421053,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.1703277587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1836,
+      "loss/crossentropy": 1.9470459461212157,
+      "loss/hidden": 3.04375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27280396595597267,
+      "loss/reg": 0.0,
+      "step": 18550
+    },
+    {
+      "epoch": 0.12210526315789473,
+      "grad_norm": 4.5625,
+      "grad_norm_var": 0.52919921875,
+      "learning_rate": 0.0001,
+      "loss": 3.1155,
+      "loss/crossentropy": 2.216350567340851,
+      "loss/hidden": 2.83125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22622163146734237,
+      "loss/reg": 0.0,
+      "step": 18560
+    },
+    {
+      "epoch": 0.12217105263157894,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.4434733072916667,
+      "learning_rate": 0.0001,
+      "loss": 3.1029,
+      "loss/crossentropy": 2.1628998279571534,
+      "loss/hidden": 3.0609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28237638175487517,
+      "loss/reg": 0.0,
+      "step": 18570
+    },
+    {
+      "epoch": 0.12223684210526316,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.3641916910807292,
+      "learning_rate": 0.0001,
+      "loss": 3.1955,
+      "loss/crossentropy": 2.486808693408966,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2668070778250694,
+      "loss/reg": 0.0,
+      "step": 18580
+    },
+    {
+      "epoch": 0.12230263157894737,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.3584950764973958,
+      "learning_rate": 0.0001,
+      "loss": 3.1489,
+      "loss/crossentropy": 2.3242167592048646,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2651012405753136,
+      "loss/reg": 0.0,
+      "step": 18590
+    },
+    {
+      "epoch": 0.12236842105263158,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.05087890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0364,
+      "loss/crossentropy": 2.4761088371276854,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24524887502193451,
+      "loss/reg": 0.0,
+      "step": 18600
+    },
+    {
+      "epoch": 0.12243421052631578,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.08817952473958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1836,
+      "loss/crossentropy": 2.012187111377716,
+      "loss/hidden": 3.0765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2920076042413712,
+      "loss/reg": 0.0,
+      "step": 18610
+    },
+    {
+      "epoch": 0.1225,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.07818094889322917,
+      "learning_rate": 0.0001,
+      "loss": 3.1844,
+      "loss/crossentropy": 2.3601612210273744,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2700245052576065,
+      "loss/reg": 0.0,
+      "step": 18620
+    },
+    {
+      "epoch": 0.12256578947368421,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.6979482014973958,
+      "learning_rate": 0.0001,
+      "loss": 3.1953,
+      "loss/crossentropy": 2.2805428981781004,
+      "loss/hidden": 2.740625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21399059891700745,
+      "loss/reg": 0.0,
+      "step": 18630
+    },
+    {
+      "epoch": 0.12263157894736842,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.09913736979166667,
+      "learning_rate": 0.0001,
+      "loss": 3.2462,
+      "loss/crossentropy": 2.3119712233543397,
+      "loss/hidden": 3.084375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27173476070165636,
+      "loss/reg": 0.0,
+      "step": 18640
+    },
+    {
+      "epoch": 0.12269736842105264,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.12066650390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1387,
+      "loss/crossentropy": 2.1647680759429933,
+      "loss/hidden": 2.9765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2955679178237915,
+      "loss/reg": 0.0,
+      "step": 18650
+    },
+    {
+      "epoch": 0.12276315789473684,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.0642242431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0672,
+      "loss/crossentropy": 2.398548412322998,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24050813913345337,
+      "loss/reg": 0.0,
+      "step": 18660
+    },
+    {
+      "epoch": 0.12282894736842105,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.43463134765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1768,
+      "loss/crossentropy": 2.007632791996002,
+      "loss/hidden": 3.2265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28355503678321836,
+      "loss/reg": 0.0,
+      "step": 18670
+    },
+    {
+      "epoch": 0.12289473684210526,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.41833394368489585,
+      "learning_rate": 0.0001,
+      "loss": 3.1491,
+      "loss/crossentropy": 2.4610511898994445,
+      "loss/hidden": 2.7359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23162013590335845,
+      "loss/reg": 0.0,
+      "step": 18680
+    },
+    {
+      "epoch": 0.12296052631578948,
+      "grad_norm": 4.15625,
+      "grad_norm_var": 0.29620335896809896,
+      "learning_rate": 0.0001,
+      "loss": 3.1762,
+      "loss/crossentropy": 2.11824317574501,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2083466961979866,
+      "loss/reg": 0.0,
+      "step": 18690
+    },
+    {
+      "epoch": 0.12302631578947368,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 1.1225685119628905,
+      "learning_rate": 0.0001,
+      "loss": 3.1766,
+      "loss/crossentropy": 2.236714720726013,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23704309910535812,
+      "loss/reg": 0.0,
+      "step": 18700
+    },
+    {
+      "epoch": 0.12309210526315789,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.12454020182291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0992,
+      "loss/crossentropy": 2.2541938424110413,
+      "loss/hidden": 2.909375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23853187412023544,
+      "loss/reg": 0.0,
+      "step": 18710
+    },
+    {
+      "epoch": 0.1231578947368421,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.070654296875,
+      "learning_rate": 0.0001,
+      "loss": 3.0403,
+      "loss/crossentropy": 2.4512630701065063,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24162033200263977,
+      "loss/reg": 0.0,
+      "step": 18720
+    },
+    {
+      "epoch": 0.12322368421052632,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.03876546223958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1522,
+      "loss/crossentropy": 2.3199268102645876,
+      "loss/hidden": 2.85625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23435179740190507,
+      "loss/reg": 0.0,
+      "step": 18730
+    },
+    {
+      "epoch": 0.12328947368421053,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.0454254150390625,
+      "learning_rate": 0.0001,
+      "loss": 3.019,
+      "loss/crossentropy": 2.3065245509147645,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22434473037719727,
+      "loss/reg": 0.0,
+      "step": 18740
+    },
+    {
+      "epoch": 0.12335526315789473,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.1038726806640625,
+      "learning_rate": 0.0001,
+      "loss": 3.145,
+      "loss/crossentropy": 2.2620524525642396,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2564009681344032,
+      "loss/reg": 0.0,
+      "step": 18750
+    },
+    {
+      "epoch": 0.12342105263157895,
+      "grad_norm": 3.40625,
+      "grad_norm_var": 0.1225982666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1243,
+      "loss/crossentropy": 2.06068754196167,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23558274507522584,
+      "loss/reg": 0.0,
+      "step": 18760
+    },
+    {
+      "epoch": 0.12348684210526316,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.28750712076822915,
+      "learning_rate": 0.0001,
+      "loss": 3.1153,
+      "loss/crossentropy": 2.5775238513946532,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2552958935499191,
+      "loss/reg": 0.0,
+      "step": 18770
+    },
+    {
+      "epoch": 0.12355263157894737,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.09669596354166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1108,
+      "loss/crossentropy": 2.100378167629242,
+      "loss/hidden": 2.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23871416002511978,
+      "loss/reg": 0.0,
+      "step": 18780
+    },
+    {
+      "epoch": 0.12361842105263159,
+      "grad_norm": 5.5,
+      "grad_norm_var": 0.6776112874348958,
+      "learning_rate": 0.0001,
+      "loss": 3.1095,
+      "loss/crossentropy": 2.4402441143989564,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23556852638721465,
+      "loss/reg": 0.0,
+      "step": 18790
+    },
+    {
+      "epoch": 0.12368421052631579,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.6309967041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1612,
+      "loss/crossentropy": 2.3034533500671386,
+      "loss/hidden": 2.8390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2525601238012314,
+      "loss/reg": 0.0,
+      "step": 18800
+    },
+    {
+      "epoch": 0.12375,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.06685282389322916,
+      "learning_rate": 0.0001,
+      "loss": 3.1566,
+      "loss/crossentropy": 2.447878336906433,
+      "loss/hidden": 2.8234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2531526446342468,
+      "loss/reg": 0.0,
+      "step": 18810
+    },
+    {
+      "epoch": 0.12381578947368421,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.17978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1701,
+      "loss/crossentropy": 2.297843897342682,
+      "loss/hidden": 2.7875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2250390335917473,
+      "loss/reg": 0.0,
+      "step": 18820
+    },
+    {
+      "epoch": 0.12388157894736843,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.06181640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1129,
+      "loss/crossentropy": 2.1577285885810853,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2582122042775154,
+      "loss/reg": 0.0,
+      "step": 18830
+    },
+    {
+      "epoch": 0.12394736842105263,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.08957926432291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1521,
+      "loss/crossentropy": 2.5041862964630126,
+      "loss/hidden": 2.840625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24232363551855088,
+      "loss/reg": 0.0,
+      "step": 18840
+    },
+    {
+      "epoch": 0.12401315789473684,
+      "grad_norm": 2.953125,
+      "grad_norm_var": 0.10745035807291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1007,
+      "loss/crossentropy": 1.8846357107162475,
+      "loss/hidden": 3.096875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2533442348241806,
+      "loss/reg": 0.0,
+      "step": 18850
+    },
+    {
+      "epoch": 0.12407894736842105,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.11793212890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0916,
+      "loss/crossentropy": 2.5479671955108643,
+      "loss/hidden": 2.8796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26022554039955137,
+      "loss/reg": 0.0,
+      "step": 18860
+    },
+    {
+      "epoch": 0.12414473684210527,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.4687652587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1202,
+      "loss/crossentropy": 2.3512953519821167,
+      "loss/hidden": 3.0296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2577581197023392,
+      "loss/reg": 0.0,
+      "step": 18870
+    },
+    {
+      "epoch": 0.12421052631578948,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.49575907389322915,
+      "learning_rate": 0.0001,
+      "loss": 3.192,
+      "loss/crossentropy": 2.4191364645957947,
+      "loss/hidden": 2.9328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2868465960025787,
+      "loss/reg": 0.0,
+      "step": 18880
+    },
+    {
+      "epoch": 0.12427631578947368,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.20038655598958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1072,
+      "loss/crossentropy": 2.548805284500122,
+      "loss/hidden": 2.6828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22707584351301194,
+      "loss/reg": 0.0,
+      "step": 18890
+    },
+    {
+      "epoch": 0.12434210526315789,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.16153055826822918,
+      "learning_rate": 0.0001,
+      "loss": 3.0653,
+      "loss/crossentropy": 2.4287894129753114,
+      "loss/hidden": 2.8359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24110280126333236,
+      "loss/reg": 0.0,
+      "step": 18900
+    },
+    {
+      "epoch": 0.1244078947368421,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.06378580729166666,
+      "learning_rate": 0.0001,
+      "loss": 3.1229,
+      "loss/crossentropy": 2.381858563423157,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28408930599689486,
+      "loss/reg": 0.0,
+      "step": 18910
+    },
+    {
+      "epoch": 0.12447368421052632,
+      "grad_norm": 3.9375,
+      "grad_norm_var": 0.33860575358072914,
+      "learning_rate": 0.0001,
+      "loss": 3.1459,
+      "loss/crossentropy": 2.0421772241592406,
+      "loss/hidden": 2.8796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24683285355567933,
+      "loss/reg": 0.0,
+      "step": 18920
+    },
+    {
+      "epoch": 0.12453947368421053,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.3823638916015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1332,
+      "loss/crossentropy": 2.528811717033386,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26049076169729235,
+      "loss/reg": 0.0,
+      "step": 18930
+    },
+    {
+      "epoch": 0.12460526315789473,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.1140045166015625,
+      "learning_rate": 0.0001,
+      "loss": 3.2262,
+      "loss/crossentropy": 2.391852283477783,
+      "loss/hidden": 3.084375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3043837010860443,
+      "loss/reg": 0.0,
+      "step": 18940
+    },
+    {
+      "epoch": 0.12467105263157895,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.06843973795572916,
+      "learning_rate": 0.0001,
+      "loss": 3.149,
+      "loss/crossentropy": 2.218567681312561,
+      "loss/hidden": 2.9609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.254660502076149,
+      "loss/reg": 0.0,
+      "step": 18950
+    },
+    {
+      "epoch": 0.12473684210526316,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.0237701416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1052,
+      "loss/crossentropy": 2.3154717803001406,
+      "loss/hidden": 2.815625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2391164407134056,
+      "loss/reg": 0.0,
+      "step": 18960
+    },
+    {
+      "epoch": 0.12480263157894737,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.05070699055989583,
+      "learning_rate": 0.0001,
+      "loss": 3.0759,
+      "loss/crossentropy": 2.2152688026428224,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28472713232040403,
+      "loss/reg": 0.0,
+      "step": 18970
+    },
+    {
+      "epoch": 0.12486842105263157,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.08042704264322917,
+      "learning_rate": 0.0001,
+      "loss": 3.1049,
+      "loss/crossentropy": 2.392467772960663,
+      "loss/hidden": 3.1921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34464606642723083,
+      "loss/reg": 0.0,
+      "step": 18980
+    },
+    {
+      "epoch": 0.12493421052631579,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.07502848307291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1321,
+      "loss/crossentropy": 2.4656677722930906,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25502916276454923,
+      "loss/reg": 0.0,
+      "step": 18990
+    },
+    {
+      "epoch": 0.125,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 1.4638824462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.2055,
+      "loss/crossentropy": 2.27457115650177,
+      "loss/hidden": 2.70625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21155266612768173,
+      "loss/reg": 0.0,
+      "step": 19000
+    },
+    {
+      "epoch": 0.1250657894736842,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.5252115885416667,
+      "learning_rate": 0.0001,
+      "loss": 3.0721,
+      "loss/crossentropy": 2.232183575630188,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23653523325920106,
+      "loss/reg": 0.0,
+      "step": 19010
+    },
+    {
+      "epoch": 0.12513157894736843,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.3392415364583333,
+      "learning_rate": 0.0001,
+      "loss": 3.1599,
+      "loss/crossentropy": 2.134384286403656,
+      "loss/hidden": 2.8203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23607225120067596,
+      "loss/reg": 0.0,
+      "step": 19020
+    },
+    {
+      "epoch": 0.12519736842105264,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.14305013020833332,
+      "learning_rate": 0.0001,
+      "loss": 3.1322,
+      "loss/crossentropy": 2.406242322921753,
+      "loss/hidden": 2.91875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2707079291343689,
+      "loss/reg": 0.0,
+      "step": 19030
+    },
+    {
+      "epoch": 0.12526315789473685,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.34036051432291664,
+      "learning_rate": 0.0001,
+      "loss": 3.2069,
+      "loss/crossentropy": 2.4997928857803347,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2432278200984001,
+      "loss/reg": 0.0,
+      "step": 19040
+    },
+    {
+      "epoch": 0.12532894736842104,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.12541910807291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0995,
+      "loss/crossentropy": 2.2165623545646667,
+      "loss/hidden": 2.78125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24738555699586867,
+      "loss/reg": 0.0,
+      "step": 19050
+    },
+    {
+      "epoch": 0.12539473684210525,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.15855712890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1697,
+      "loss/crossentropy": 2.510625755786896,
+      "loss/hidden": 2.7984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22799582332372664,
+      "loss/reg": 0.0,
+      "step": 19060
+    },
+    {
+      "epoch": 0.12546052631578947,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.09364827473958333,
+      "learning_rate": 0.0001,
+      "loss": 3.2057,
+      "loss/crossentropy": 2.380051004886627,
+      "loss/hidden": 2.95,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2702123373746872,
+      "loss/reg": 0.0,
+      "step": 19070
+    },
+    {
+      "epoch": 0.12552631578947368,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.07896728515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0865,
+      "loss/crossentropy": 2.2913742661476135,
+      "loss/hidden": 2.740625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22653487473726272,
+      "loss/reg": 0.0,
+      "step": 19080
+    },
+    {
+      "epoch": 0.1255921052631579,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.1556549072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.1844,
+      "loss/crossentropy": 2.39451003074646,
+      "loss/hidden": 2.909375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23494229055941104,
+      "loss/reg": 0.0,
+      "step": 19090
+    },
+    {
+      "epoch": 0.1256578947368421,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.42444254557291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1173,
+      "loss/crossentropy": 2.2207372069358824,
+      "loss/hidden": 2.8984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2351771742105484,
+      "loss/reg": 0.0,
+      "step": 19100
+    },
+    {
+      "epoch": 0.12572368421052632,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.405517578125,
+      "learning_rate": 0.0001,
+      "loss": 3.1686,
+      "loss/crossentropy": 2.544073963165283,
+      "loss/hidden": 3.00625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23578422516584396,
+      "loss/reg": 0.0,
+      "step": 19110
+    },
+    {
+      "epoch": 0.12578947368421053,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.022623697916666668,
+      "learning_rate": 0.0001,
+      "loss": 3.1575,
+      "loss/crossentropy": 2.0650166511535644,
+      "loss/hidden": 2.9515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2400606468319893,
+      "loss/reg": 0.0,
+      "step": 19120
+    },
+    {
+      "epoch": 0.12585526315789475,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.12542215983072916,
+      "learning_rate": 0.0001,
+      "loss": 3.1717,
+      "loss/crossentropy": 2.1879891753196716,
+      "loss/hidden": 2.8234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23869821876287461,
+      "loss/reg": 0.0,
+      "step": 19130
+    },
+    {
+      "epoch": 0.12592105263157893,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 2.683204189608586e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2235,
+      "loss/crossentropy": 2.4875245571136473,
+      "loss/hidden": 2.9796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3171342611312866,
+      "loss/reg": 0.0,
+      "step": 19140
+    },
+    {
+      "epoch": 0.12598684210526315,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.12364908854166666,
+      "learning_rate": 0.0001,
+      "loss": 3.1832,
+      "loss/crossentropy": 2.199462330341339,
+      "loss/hidden": 3.1125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2820486217737198,
+      "loss/reg": 0.0,
+      "step": 19150
+    },
+    {
+      "epoch": 0.12605263157894736,
+      "grad_norm": 3.53125,
+      "grad_norm_var": 0.133642578125,
+      "learning_rate": 0.0001,
+      "loss": 3.2179,
+      "loss/crossentropy": 2.433344876766205,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3452912583947182,
+      "loss/reg": 0.0,
+      "step": 19160
+    },
+    {
+      "epoch": 0.12611842105263157,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.23087946573893228,
+      "learning_rate": 0.0001,
+      "loss": 3.073,
+      "loss/crossentropy": 2.107306253910065,
+      "loss/hidden": 3.0140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29124595075845716,
+      "loss/reg": 0.0,
+      "step": 19170
+    },
+    {
+      "epoch": 0.1261842105263158,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 2.206763811704668e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2818,
+      "loss/crossentropy": 2.17460697889328,
+      "loss/hidden": 2.8359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26940477788448336,
+      "loss/reg": 0.0,
+      "step": 19180
+    },
+    {
+      "epoch": 0.12625,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.06741434733072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1977,
+      "loss/crossentropy": 2.249357485771179,
+      "loss/hidden": 2.95625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29158340096473695,
+      "loss/reg": 0.0,
+      "step": 19190
+    },
+    {
+      "epoch": 0.12631578947368421,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.1665679931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0798,
+      "loss/crossentropy": 2.39562349319458,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24128414690494537,
+      "loss/reg": 0.0,
+      "step": 19200
+    },
+    {
+      "epoch": 0.12638157894736843,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.04352925618489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0886,
+      "loss/crossentropy": 2.4379887223243712,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2693786233663559,
+      "loss/reg": 0.0,
+      "step": 19210
+    },
+    {
+      "epoch": 0.12644736842105264,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.05164286295572917,
+      "learning_rate": 0.0001,
+      "loss": 3.052,
+      "loss/crossentropy": 2.327665722370148,
+      "loss/hidden": 2.8140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.37138707339763644,
+      "loss/reg": 0.0,
+      "step": 19220
+    },
+    {
+      "epoch": 0.12651315789473686,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 1.5772623697916666,
+      "learning_rate": 0.0001,
+      "loss": 3.1007,
+      "loss/crossentropy": 2.1090051174163817,
+      "loss/hidden": 3.0875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23532682955265044,
+      "loss/reg": 0.0,
+      "step": 19230
+    },
+    {
+      "epoch": 0.12657894736842104,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.76259765625,
+      "learning_rate": 0.0001,
+      "loss": 3.087,
+      "loss/crossentropy": 2.3096412897109984,
+      "loss/hidden": 2.8953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2415456846356392,
+      "loss/reg": 0.0,
+      "step": 19240
+    },
+    {
+      "epoch": 0.12664473684210525,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.0485504150390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0857,
+      "loss/crossentropy": 2.252851128578186,
+      "loss/hidden": 2.7515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21757592558860778,
+      "loss/reg": 0.0,
+      "step": 19250
+    },
+    {
+      "epoch": 0.12671052631578947,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.16155192057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1508,
+      "loss/crossentropy": 2.19182807803154,
+      "loss/hidden": 2.940625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23997026532888413,
+      "loss/reg": 0.0,
+      "step": 19260
+    },
+    {
+      "epoch": 0.12677631578947368,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.4118609110514323,
+      "learning_rate": 0.0001,
+      "loss": 3.1121,
+      "loss/crossentropy": 2.4667071104049683,
+      "loss/hidden": 2.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23552417606115342,
+      "loss/reg": 0.0,
+      "step": 19270
+    },
+    {
+      "epoch": 0.1268421052631579,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.04609553019205729,
+      "learning_rate": 0.0001,
+      "loss": 3.1763,
+      "loss/crossentropy": 2.1369692206382753,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24942600578069687,
+      "loss/reg": 0.0,
+      "step": 19280
+    },
+    {
+      "epoch": 0.1269078947368421,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.0526031494140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1068,
+      "loss/crossentropy": 2.11109459400177,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25110483914613724,
+      "loss/reg": 0.0,
+      "step": 19290
+    },
+    {
+      "epoch": 0.12697368421052632,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.053343709309895834,
+      "learning_rate": 0.0001,
+      "loss": 3.1517,
+      "loss/crossentropy": 2.380664014816284,
+      "loss/hidden": 2.8765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2508363798260689,
+      "loss/reg": 0.0,
+      "step": 19300
+    },
+    {
+      "epoch": 0.12703947368421054,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.034032185872395836,
+      "learning_rate": 0.0001,
+      "loss": 3.1303,
+      "loss/crossentropy": 2.380405902862549,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2564548909664154,
+      "loss/reg": 0.0,
+      "step": 19310
+    },
+    {
+      "epoch": 0.12710526315789475,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.03234049479166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1223,
+      "loss/crossentropy": 2.1872188091278075,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2912978962063789,
+      "loss/reg": 0.0,
+      "step": 19320
+    },
+    {
+      "epoch": 0.12717105263157893,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.07059504191080729,
+      "learning_rate": 0.0001,
+      "loss": 3.1221,
+      "loss/crossentropy": 2.470194971561432,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2459412097930908,
+      "loss/reg": 0.0,
+      "step": 19330
+    },
+    {
+      "epoch": 0.12723684210526315,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.1386431376139323,
+      "learning_rate": 0.0001,
+      "loss": 3.1555,
+      "loss/crossentropy": 2.3521093368530273,
+      "loss/hidden": 2.925,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2635854005813599,
+      "loss/reg": 0.0,
+      "step": 19340
+    },
+    {
+      "epoch": 0.12730263157894736,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.053857421875,
+      "learning_rate": 0.0001,
+      "loss": 3.0594,
+      "loss/crossentropy": 2.099438285827637,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22058220505714415,
+      "loss/reg": 0.0,
+      "step": 19350
+    },
+    {
+      "epoch": 0.12736842105263158,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.10086161295572917,
+      "learning_rate": 0.0001,
+      "loss": 3.1877,
+      "loss/crossentropy": 2.2927380204200745,
+      "loss/hidden": 2.8203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25694535821676256,
+      "loss/reg": 0.0,
+      "step": 19360
+    },
+    {
+      "epoch": 0.1274342105263158,
+      "grad_norm": 1.9375,
+      "grad_norm_var": 0.05416259765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0826,
+      "loss/crossentropy": 2.1295772194862366,
+      "loss/hidden": 3.0015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.254076661169529,
+      "loss/reg": 0.0,
+      "step": 19370
+    },
+    {
+      "epoch": 0.1275,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.24744466145833333,
+      "learning_rate": 0.0001,
+      "loss": 3.1415,
+      "loss/crossentropy": 1.9256490916013718,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24017905220389366,
+      "loss/reg": 0.0,
+      "step": 19380
+    },
+    {
+      "epoch": 0.12756578947368422,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.5484659830729167,
+      "learning_rate": 0.0001,
+      "loss": 3.1094,
+      "loss/crossentropy": 2.4685181736946107,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26331629455089567,
+      "loss/reg": 0.0,
+      "step": 19390
+    },
+    {
+      "epoch": 0.12763157894736843,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.39381103515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1037,
+      "loss/crossentropy": 2.2764881372451784,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25390326231718063,
+      "loss/reg": 0.0,
+      "step": 19400
+    },
+    {
+      "epoch": 0.12769736842105264,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.0698150634765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1262,
+      "loss/crossentropy": 2.299801528453827,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2553006038069725,
+      "loss/reg": 0.0,
+      "step": 19410
+    },
+    {
+      "epoch": 0.12776315789473683,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.266357421875,
+      "learning_rate": 0.0001,
+      "loss": 3.1239,
+      "loss/crossentropy": 2.0346228003501894,
+      "loss/hidden": 2.9515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22659170776605606,
+      "loss/reg": 0.0,
+      "step": 19420
+    },
+    {
+      "epoch": 0.12782894736842104,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.37324930826822916,
+      "learning_rate": 0.0001,
+      "loss": 3.1527,
+      "loss/crossentropy": 1.8951176881790162,
+      "loss/hidden": 2.83125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23443188220262529,
+      "loss/reg": 0.0,
+      "step": 19430
+    },
+    {
+      "epoch": 0.12789473684210526,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.30020243326822915,
+      "learning_rate": 0.0001,
+      "loss": 3.1742,
+      "loss/crossentropy": 2.0540109515190124,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20190905332565307,
+      "loss/reg": 0.0,
+      "step": 19440
+    },
+    {
+      "epoch": 0.12796052631578947,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.19728190104166668,
+      "learning_rate": 0.0001,
+      "loss": 3.1528,
+      "loss/crossentropy": 2.3164134502410887,
+      "loss/hidden": 2.8234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2684441477060318,
+      "loss/reg": 0.0,
+      "step": 19450
+    },
+    {
+      "epoch": 0.12802631578947368,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.2419586181640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1599,
+      "loss/crossentropy": 2.380271017551422,
+      "loss/hidden": 2.96875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2657022625207901,
+      "loss/reg": 0.0,
+      "step": 19460
+    },
+    {
+      "epoch": 0.1280921052631579,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.11366780598958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1441,
+      "loss/crossentropy": 2.4884063005447388,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2683678835630417,
+      "loss/reg": 0.0,
+      "step": 19470
+    },
+    {
+      "epoch": 0.1281578947368421,
+      "grad_norm": 2038431744.0,
+      "grad_norm_var": 2.597002478369833e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2142,
+      "loss/crossentropy": 2.0524453282356263,
+      "loss/hidden": 3.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2409697949886322,
+      "loss/reg": 0.0,
+      "step": 19480
+    },
+    {
+      "epoch": 0.12822368421052632,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 2.597002478497235e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0528,
+      "loss/crossentropy": 2.2681432604789733,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24559762477874755,
+      "loss/reg": 0.0,
+      "step": 19490
+    },
+    {
+      "epoch": 0.12828947368421054,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.030345662434895834,
+      "learning_rate": 0.0001,
+      "loss": 3.0796,
+      "loss/crossentropy": 2.362069141864777,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2367064341902733,
+      "loss/reg": 0.0,
+      "step": 19500
+    },
+    {
+      "epoch": 0.12835526315789475,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.059382120768229164,
+      "learning_rate": 0.0001,
+      "loss": 3.1465,
+      "loss/crossentropy": 2.260143554210663,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25476695597171783,
+      "loss/reg": 0.0,
+      "step": 19510
+    },
+    {
+      "epoch": 0.12842105263157894,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.18327534993489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1443,
+      "loss/crossentropy": 2.322359097003937,
+      "loss/hidden": 2.7875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22897413671016692,
+      "loss/reg": 0.0,
+      "step": 19520
+    },
+    {
+      "epoch": 0.12848684210526315,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.4216379801432292,
+      "learning_rate": 0.0001,
+      "loss": 3.1736,
+      "loss/crossentropy": 2.658802056312561,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26290144920349123,
+      "loss/reg": 0.0,
+      "step": 19530
+    },
+    {
+      "epoch": 0.12855263157894736,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.1238677978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1818,
+      "loss/crossentropy": 2.389211559295654,
+      "loss/hidden": 2.8796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29232275635004046,
+      "loss/reg": 0.0,
+      "step": 19540
+    },
+    {
+      "epoch": 0.12861842105263158,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.049845123291015626,
+      "learning_rate": 0.0001,
+      "loss": 3.0708,
+      "loss/crossentropy": 2.186306023597717,
+      "loss/hidden": 2.99375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2743732765316963,
+      "loss/reg": 0.0,
+      "step": 19550
+    },
+    {
+      "epoch": 0.1286842105263158,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.024887847900390624,
+      "learning_rate": 0.0001,
+      "loss": 3.1515,
+      "loss/crossentropy": 2.325215721130371,
+      "loss/hidden": 2.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24441724121570588,
+      "loss/reg": 0.0,
+      "step": 19560
+    },
+    {
+      "epoch": 0.12875,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.18791910807291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1163,
+      "loss/crossentropy": 2.4399210453033446,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23570073395967484,
+      "loss/reg": 0.0,
+      "step": 19570
+    },
+    {
+      "epoch": 0.12881578947368422,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.0600250244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0731,
+      "loss/crossentropy": 2.250354325771332,
+      "loss/hidden": 2.9390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25124100893735885,
+      "loss/reg": 0.0,
+      "step": 19580
+    },
+    {
+      "epoch": 0.12888157894736843,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.2105865478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1494,
+      "loss/crossentropy": 2.3004735589027403,
+      "loss/hidden": 2.93125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.250587160885334,
+      "loss/reg": 0.0,
+      "step": 19590
+    },
+    {
+      "epoch": 0.12894736842105264,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.23645833333333333,
+      "learning_rate": 0.0001,
+      "loss": 3.099,
+      "loss/crossentropy": 2.3559111833572386,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26171091198921204,
+      "loss/reg": 0.0,
+      "step": 19600
+    },
+    {
+      "epoch": 0.12901315789473683,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.028955078125,
+      "learning_rate": 0.0001,
+      "loss": 3.171,
+      "loss/crossentropy": 2.2754984378814695,
+      "loss/hidden": 3.046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2695039168000221,
+      "loss/reg": 0.0,
+      "step": 19610
+    },
+    {
+      "epoch": 0.12907894736842104,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.12908426920572916,
+      "learning_rate": 0.0001,
+      "loss": 3.1609,
+      "loss/crossentropy": 2.457037115097046,
+      "loss/hidden": 2.8,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2523172840476036,
+      "loss/reg": 0.0,
+      "step": 19620
+    },
+    {
+      "epoch": 0.12914473684210526,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.14287007649739583,
+      "learning_rate": 0.0001,
+      "loss": 3.0694,
+      "loss/crossentropy": 2.33393777012825,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23025162070989608,
+      "loss/reg": 0.0,
+      "step": 19630
+    },
+    {
+      "epoch": 0.12921052631578947,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.029618326822916666,
+      "learning_rate": 0.0001,
+      "loss": 3.0684,
+      "loss/crossentropy": 2.155827796459198,
+      "loss/hidden": 2.8140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22635919600725174,
+      "loss/reg": 0.0,
+      "step": 19640
+    },
+    {
+      "epoch": 0.12927631578947368,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.31851806640625,
+      "learning_rate": 0.0001,
+      "loss": 3.112,
+      "loss/crossentropy": 2.2350030899047852,
+      "loss/hidden": 2.9109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2474326401948929,
+      "loss/reg": 0.0,
+      "step": 19650
+    },
+    {
+      "epoch": 0.1293421052631579,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.30481363932291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0884,
+      "loss/crossentropy": 2.1235520601272584,
+      "loss/hidden": 2.84375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23340977281332015,
+      "loss/reg": 0.0,
+      "step": 19660
+    },
+    {
+      "epoch": 0.1294078947368421,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.058958943684895834,
+      "learning_rate": 0.0001,
+      "loss": 3.1092,
+      "loss/crossentropy": 2.436754751205444,
+      "loss/hidden": 2.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2396928071975708,
+      "loss/reg": 0.0,
+      "step": 19670
+    },
+    {
+      "epoch": 0.12947368421052632,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.2934804280598958,
+      "learning_rate": 0.0001,
+      "loss": 3.0531,
+      "loss/crossentropy": 2.24703209400177,
+      "loss/hidden": 2.8359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23143238127231597,
+      "loss/reg": 0.0,
+      "step": 19680
+    },
+    {
+      "epoch": 0.12953947368421054,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.25420303344726564,
+      "learning_rate": 0.0001,
+      "loss": 3.0318,
+      "loss/crossentropy": 2.536008381843567,
+      "loss/hidden": 2.7078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22230196446180345,
+      "loss/reg": 0.0,
+      "step": 19690
+    },
+    {
+      "epoch": 0.12960526315789472,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.29778416951497394,
+      "learning_rate": 0.0001,
+      "loss": 3.0902,
+      "loss/crossentropy": 2.1229737393558024,
+      "loss/hidden": 2.821875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2103697349317372,
+      "loss/reg": 0.0,
+      "step": 19700
+    },
+    {
+      "epoch": 0.12967105263157894,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.09654032389322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0303,
+      "loss/crossentropy": 2.360739004611969,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.238627889752388,
+      "loss/reg": 0.0,
+      "step": 19710
+    },
+    {
+      "epoch": 0.12973684210526315,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.16044514973958332,
+      "learning_rate": 0.0001,
+      "loss": 3.082,
+      "loss/crossentropy": 2.3643787026405336,
+      "loss/hidden": 2.9796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2354188710451126,
+      "loss/reg": 0.0,
+      "step": 19720
+    },
+    {
+      "epoch": 0.12980263157894736,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.23007405598958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0884,
+      "loss/crossentropy": 2.3766650319099427,
+      "loss/hidden": 3.0109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26477697044610976,
+      "loss/reg": 0.0,
+      "step": 19730
+    },
+    {
+      "epoch": 0.12986842105263158,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.055946604410807295,
+      "learning_rate": 0.0001,
+      "loss": 3.0414,
+      "loss/crossentropy": 2.6002285480499268,
+      "loss/hidden": 2.98125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33312758058309555,
+      "loss/reg": 0.0,
+      "step": 19740
+    },
+    {
+      "epoch": 0.1299342105263158,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.13170547485351564,
+      "learning_rate": 0.0001,
+      "loss": 3.0758,
+      "loss/crossentropy": 2.457702159881592,
+      "loss/hidden": 2.9890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25578114166855814,
+      "loss/reg": 0.0,
+      "step": 19750
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.03232014973958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0882,
+      "loss/crossentropy": 2.261428934335709,
+      "loss/hidden": 2.753125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2278106167912483,
+      "loss/reg": 0.0,
+      "step": 19760
+    },
+    {
+      "epoch": 0.13006578947368422,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.022248331705729166,
+      "learning_rate": 0.0001,
+      "loss": 3.0582,
+      "loss/crossentropy": 2.060434710979462,
+      "loss/hidden": 2.9046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24812956005334855,
+      "loss/reg": 0.0,
+      "step": 19770
+    },
+    {
+      "epoch": 0.13013157894736843,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.0256256103515625,
+      "learning_rate": 0.0001,
+      "loss": 3.08,
+      "loss/crossentropy": 2.2472903966903686,
+      "loss/hidden": 3.1,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34347400814294815,
+      "loss/reg": 0.0,
+      "step": 19780
+    },
+    {
+      "epoch": 0.13019736842105264,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.0520660400390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1154,
+      "loss/crossentropy": 2.3003188014030456,
+      "loss/hidden": 2.9625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2695119082927704,
+      "loss/reg": 0.0,
+      "step": 19790
+    },
+    {
+      "epoch": 0.13026315789473683,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.08396809895833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0827,
+      "loss/crossentropy": 2.2551465153694155,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24707757085561752,
+      "loss/reg": 0.0,
+      "step": 19800
+    },
+    {
+      "epoch": 0.13032894736842104,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.15972900390625,
+      "learning_rate": 0.0001,
+      "loss": 3.151,
+      "loss/crossentropy": 2.323357033729553,
+      "loss/hidden": 2.89375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2660538278520107,
+      "loss/reg": 0.0,
+      "step": 19810
+    },
+    {
+      "epoch": 0.13039473684210526,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.15328776041666667,
+      "learning_rate": 0.0001,
+      "loss": 3.0792,
+      "loss/crossentropy": 2.369428300857544,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2439693480730057,
+      "loss/reg": 0.0,
+      "step": 19820
+    },
+    {
+      "epoch": 0.13046052631578947,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.09102274576822916,
+      "learning_rate": 0.0001,
+      "loss": 3.1636,
+      "loss/crossentropy": 2.2469497442245485,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24479606077075006,
+      "loss/reg": 0.0,
+      "step": 19830
+    },
+    {
+      "epoch": 0.13052631578947368,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.052302042643229164,
+      "learning_rate": 0.0001,
+      "loss": 3.1694,
+      "loss/crossentropy": 2.4753658294677736,
+      "loss/hidden": 3.0390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3045656159520149,
+      "loss/reg": 0.0,
+      "step": 19840
+    },
+    {
+      "epoch": 0.1305921052631579,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.0764068603515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0959,
+      "loss/crossentropy": 2.4656317472457885,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.258389513194561,
+      "loss/reg": 0.0,
+      "step": 19850
+    },
+    {
+      "epoch": 0.1306578947368421,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.06215718587239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0406,
+      "loss/crossentropy": 2.440659189224243,
+      "loss/hidden": 2.7171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26817646920681,
+      "loss/reg": 0.0,
+      "step": 19860
+    },
+    {
+      "epoch": 0.13072368421052633,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.04420166015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0968,
+      "loss/crossentropy": 2.206334137916565,
+      "loss/hidden": 2.8671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24976521283388137,
+      "loss/reg": 0.0,
+      "step": 19870
+    },
+    {
+      "epoch": 0.13078947368421054,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.08323567708333333,
+      "learning_rate": 0.0001,
+      "loss": 3.0685,
+      "loss/crossentropy": 2.21451940536499,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21932003498077393,
+      "loss/reg": 0.0,
+      "step": 19880
+    },
+    {
+      "epoch": 0.13085526315789472,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.17522786458333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0628,
+      "loss/crossentropy": 2.1315925240516664,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23312209993600846,
+      "loss/reg": 0.0,
+      "step": 19890
+    },
+    {
+      "epoch": 0.13092105263157894,
+      "grad_norm": 6.28125,
+      "grad_norm_var": 0.9668528238932291,
+      "learning_rate": 0.0001,
+      "loss": 3.1207,
+      "loss/crossentropy": 2.5357746481895447,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2542704403400421,
+      "loss/reg": 0.0,
+      "step": 19900
+    },
+    {
+      "epoch": 0.13098684210526315,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.9971913655598958,
+      "learning_rate": 0.0001,
+      "loss": 3.0247,
+      "loss/crossentropy": 2.53050377368927,
+      "loss/hidden": 2.6515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2318735808134079,
+      "loss/reg": 0.0,
+      "step": 19910
+    },
+    {
+      "epoch": 0.13105263157894737,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.14268290201822917,
+      "learning_rate": 0.0001,
+      "loss": 3.1545,
+      "loss/crossentropy": 2.475492477416992,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29062798619270325,
+      "loss/reg": 0.0,
+      "step": 19920
+    },
+    {
+      "epoch": 0.13111842105263158,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.08727925618489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1316,
+      "loss/crossentropy": 2.4628885269165037,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24764630049467087,
+      "loss/reg": 0.0,
+      "step": 19930
+    },
+    {
+      "epoch": 0.1311842105263158,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.07839736938476563,
+      "learning_rate": 0.0001,
+      "loss": 3.0597,
+      "loss/crossentropy": 2.46280722618103,
+      "loss/hidden": 2.6515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2163504734635353,
+      "loss/reg": 0.0,
+      "step": 19940
+    },
+    {
+      "epoch": 0.13125,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.0990875244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1648,
+      "loss/crossentropy": 2.4153407394886015,
+      "loss/hidden": 3.0875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26917385756969453,
+      "loss/reg": 0.0,
+      "step": 19950
+    },
+    {
+      "epoch": 0.13131578947368422,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.18907877604166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1173,
+      "loss/crossentropy": 2.232962656021118,
+      "loss/hidden": 3.0734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3076439991593361,
+      "loss/reg": 0.0,
+      "step": 19960
+    },
+    {
+      "epoch": 0.13138157894736843,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.13981119791666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0478,
+      "loss/crossentropy": 2.353410243988037,
+      "loss/hidden": 2.8796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24693673849105835,
+      "loss/reg": 0.0,
+      "step": 19970
+    },
+    {
+      "epoch": 0.13144736842105262,
+      "grad_norm": 3.15625,
+      "grad_norm_var": 0.1372711181640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1452,
+      "loss/crossentropy": 2.121303880214691,
+      "loss/hidden": 3.125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31650226265192033,
+      "loss/reg": 0.0,
+      "step": 19980
+    },
+    {
+      "epoch": 0.13151315789473683,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.1741607666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1028,
+      "loss/crossentropy": 2.43160719871521,
+      "loss/hidden": 2.6828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2277207463979721,
+      "loss/reg": 0.0,
+      "step": 19990
+    },
+    {
+      "epoch": 0.13157894736842105,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.15798238118489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0532,
+      "loss/crossentropy": 2.2883418917655947,
+      "loss/hidden": 2.990625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2525527849793434,
+      "loss/reg": 0.0,
+      "step": 20000
+    },
+    {
+      "epoch": 0.13164473684210526,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.10852457682291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0348,
+      "loss/crossentropy": 2.369931137561798,
+      "loss/hidden": 2.703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22242200672626494,
+      "loss/reg": 0.0,
+      "step": 20010
+    },
+    {
+      "epoch": 0.13171052631578947,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.0374664306640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0316,
+      "loss/crossentropy": 2.1939921617507934,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.234733884036541,
+      "loss/reg": 0.0,
+      "step": 20020
+    },
+    {
+      "epoch": 0.13177631578947369,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.07238667805989583,
+      "learning_rate": 0.0001,
+      "loss": 3.1278,
+      "loss/crossentropy": 2.4681461334228514,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30574188083410264,
+      "loss/reg": 0.0,
+      "step": 20030
+    },
+    {
+      "epoch": 0.1318421052631579,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.07011311848958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0111,
+      "loss/crossentropy": 2.699459183216095,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27421810775995253,
+      "loss/reg": 0.0,
+      "step": 20040
+    },
+    {
+      "epoch": 0.1319078947368421,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 41.75862528483073,
+      "learning_rate": 0.0001,
+      "loss": 3.202,
+      "loss/crossentropy": 2.263700020313263,
+      "loss/hidden": 2.84375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23062770962715148,
+      "loss/reg": 0.0,
+      "step": 20050
+    },
+    {
+      "epoch": 0.13197368421052633,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.08103841145833333,
+      "learning_rate": 0.0001,
+      "loss": 3.1056,
+      "loss/crossentropy": 2.4495514154434206,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2597358673810959,
+      "loss/reg": 0.0,
+      "step": 20060
+    },
+    {
+      "epoch": 0.13203947368421054,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.10988667805989584,
+      "learning_rate": 0.0001,
+      "loss": 3.1429,
+      "loss/crossentropy": 2.5027857065200805,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23284846246242524,
+      "loss/reg": 0.0,
+      "step": 20070
+    },
+    {
+      "epoch": 0.13210526315789473,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.4439442952473958,
+      "learning_rate": 0.0001,
+      "loss": 3.1051,
+      "loss/crossentropy": 2.3987682700157165,
+      "loss/hidden": 2.821875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2284790888428688,
+      "loss/reg": 0.0,
+      "step": 20080
+    },
+    {
+      "epoch": 0.13217105263157894,
+      "grad_norm": 4.9375,
+      "grad_norm_var": 0.8553456624348958,
+      "learning_rate": 0.0001,
+      "loss": 3.1699,
+      "loss/crossentropy": 2.415048587322235,
+      "loss/hidden": 2.9015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2737443670630455,
+      "loss/reg": 0.0,
+      "step": 20090
+    },
+    {
+      "epoch": 0.13223684210526315,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.5597005208333333,
+      "learning_rate": 0.0001,
+      "loss": 3.1123,
+      "loss/crossentropy": 2.42084618806839,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26620708405971527,
+      "loss/reg": 0.0,
+      "step": 20100
+    },
+    {
+      "epoch": 0.13230263157894737,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.022086588541666667,
+      "learning_rate": 0.0001,
+      "loss": 3.0069,
+      "loss/crossentropy": 2.4669047951698304,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22731948494911194,
+      "loss/reg": 0.0,
+      "step": 20110
+    },
+    {
+      "epoch": 0.13236842105263158,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.6456451416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1966,
+      "loss/crossentropy": 2.344004726409912,
+      "loss/hidden": 2.725,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2402897983789444,
+      "loss/reg": 0.0,
+      "step": 20120
+    },
+    {
+      "epoch": 0.1324342105263158,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 1.062083943684896,
+      "learning_rate": 0.0001,
+      "loss": 3.1146,
+      "loss/crossentropy": 2.375274932384491,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22684457302093505,
+      "loss/reg": 0.0,
+      "step": 20130
+    },
+    {
+      "epoch": 0.1325,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.046076456705729164,
+      "learning_rate": 0.0001,
+      "loss": 3.0369,
+      "loss/crossentropy": 2.366211712360382,
+      "loss/hidden": 2.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24463532716035843,
+      "loss/reg": 0.0,
+      "step": 20140
+    },
+    {
+      "epoch": 0.13256578947368422,
+      "grad_norm": 1.9375,
+      "grad_norm_var": 0.12783915201822918,
+      "learning_rate": 0.0001,
+      "loss": 3.1105,
+      "loss/crossentropy": 2.384170186519623,
+      "loss/hidden": 2.8671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28617204576730726,
+      "loss/reg": 0.0,
+      "step": 20150
+    },
+    {
+      "epoch": 0.13263157894736843,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.5388671875,
+      "learning_rate": 0.0001,
+      "loss": 3.13,
+      "loss/crossentropy": 2.37620815038681,
+      "loss/hidden": 2.7140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22599491178989412,
+      "loss/reg": 0.0,
+      "step": 20160
+    },
+    {
+      "epoch": 0.13269736842105262,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.5671875,
+      "learning_rate": 0.0001,
+      "loss": 3.0816,
+      "loss/crossentropy": 2.3014034748077394,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21943795531988144,
+      "loss/reg": 0.0,
+      "step": 20170
+    },
+    {
+      "epoch": 0.13276315789473683,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.11660868326822917,
+      "learning_rate": 0.0001,
+      "loss": 3.1416,
+      "loss/crossentropy": 2.619466185569763,
+      "loss/hidden": 3.190625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3190195769071579,
+      "loss/reg": 0.0,
+      "step": 20180
+    },
+    {
+      "epoch": 0.13282894736842105,
+      "grad_norm": 3.296875,
+      "grad_norm_var": 0.15591532389322918,
+      "learning_rate": 0.0001,
+      "loss": 3.128,
+      "loss/crossentropy": 2.3887166023254394,
+      "loss/hidden": 2.8625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27821466475725176,
+      "loss/reg": 0.0,
+      "step": 20190
+    },
+    {
+      "epoch": 0.13289473684210526,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.6188639322916667,
+      "learning_rate": 0.0001,
+      "loss": 3.0968,
+      "loss/crossentropy": 2.49861718416214,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2538172617554665,
+      "loss/reg": 0.0,
+      "step": 20200
+    },
+    {
+      "epoch": 0.13296052631578947,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.20885416666666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1671,
+      "loss/crossentropy": 2.220921754837036,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23103244453668595,
+      "loss/reg": 0.0,
+      "step": 20210
+    },
+    {
+      "epoch": 0.1330263157894737,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.18879801432291668,
+      "learning_rate": 0.0001,
+      "loss": 3.1111,
+      "loss/crossentropy": 2.333580756187439,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2729429990053177,
+      "loss/reg": 0.0,
+      "step": 20220
+    },
+    {
+      "epoch": 0.1330921052631579,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.0948883056640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0829,
+      "loss/crossentropy": 2.3158997893333435,
+      "loss/hidden": 2.6203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19884210526943208,
+      "loss/reg": 0.0,
+      "step": 20230
+    },
+    {
+      "epoch": 0.13315789473684211,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.20250244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0896,
+      "loss/crossentropy": 2.369309663772583,
+      "loss/hidden": 2.89375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26042123287916186,
+      "loss/reg": 0.0,
+      "step": 20240
+    },
+    {
+      "epoch": 0.13322368421052633,
+      "grad_norm": 3.015625,
+      "grad_norm_var": 0.17239176432291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1863,
+      "loss/crossentropy": 2.2114102065563204,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21532374620437622,
+      "loss/reg": 0.0,
+      "step": 20250
+    },
+    {
+      "epoch": 0.1332894736842105,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.33000869750976564,
+      "learning_rate": 0.0001,
+      "loss": 3.0896,
+      "loss/crossentropy": 2.4173696994781495,
+      "loss/hidden": 2.7359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2683261051774025,
+      "loss/reg": 0.0,
+      "step": 20260
+    },
+    {
+      "epoch": 0.13335526315789473,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.10729548136393229,
+      "learning_rate": 0.0001,
+      "loss": 3.1426,
+      "loss/crossentropy": 2.3808977246284484,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3049054339528084,
+      "loss/reg": 0.0,
+      "step": 20270
+    },
+    {
+      "epoch": 0.13342105263157894,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.1077789306640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1685,
+      "loss/crossentropy": 2.244146704673767,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2616411089897156,
+      "loss/reg": 0.0,
+      "step": 20280
+    },
+    {
+      "epoch": 0.13348684210526315,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.1076171875,
+      "learning_rate": 0.0001,
+      "loss": 3.1354,
+      "loss/crossentropy": 2.145916444063187,
+      "loss/hidden": 2.7375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21995762139558792,
+      "loss/reg": 0.0,
+      "step": 20290
+    },
+    {
+      "epoch": 0.13355263157894737,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.08604227701822917,
+      "learning_rate": 0.0001,
+      "loss": 3.0957,
+      "loss/crossentropy": 2.5188711285591125,
+      "loss/hidden": 2.94375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.264768448472023,
+      "loss/reg": 0.0,
+      "step": 20300
+    },
+    {
+      "epoch": 0.13361842105263158,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.3717274983723958,
+      "learning_rate": 0.0001,
+      "loss": 3.0886,
+      "loss/crossentropy": 2.392267715930939,
+      "loss/hidden": 3.1953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3307821795344353,
+      "loss/reg": 0.0,
+      "step": 20310
+    },
+    {
+      "epoch": 0.1336842105263158,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.39180399576822916,
+      "learning_rate": 0.0001,
+      "loss": 3.1034,
+      "loss/crossentropy": 2.2370328307151794,
+      "loss/hidden": 2.725,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20846432447433472,
+      "loss/reg": 0.0,
+      "step": 20320
+    },
+    {
+      "epoch": 0.13375,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.136962890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1123,
+      "loss/crossentropy": 2.358556866645813,
+      "loss/hidden": 2.8453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.264502127468586,
+      "loss/reg": 0.0,
+      "step": 20330
+    },
+    {
+      "epoch": 0.13381578947368422,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.04733072916666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1698,
+      "loss/crossentropy": 2.302602219581604,
+      "loss/hidden": 3.003125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2898910105228424,
+      "loss/reg": 0.0,
+      "step": 20340
+    },
+    {
+      "epoch": 0.13388157894736843,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.07590230305989583,
+      "learning_rate": 0.0001,
+      "loss": 3.155,
+      "loss/crossentropy": 2.2594828605651855,
+      "loss/hidden": 2.86875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2617498949170113,
+      "loss/reg": 0.0,
+      "step": 20350
+    },
+    {
+      "epoch": 0.13394736842105262,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.1132476806640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1443,
+      "loss/crossentropy": 2.6042701482772825,
+      "loss/hidden": 2.95,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2813595399260521,
+      "loss/reg": 0.0,
+      "step": 20360
+    },
+    {
+      "epoch": 0.13401315789473683,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 3.111881782679394e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1969,
+      "loss/crossentropy": 2.596519351005554,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24256999790668488,
+      "loss/reg": 0.0,
+      "step": 20370
+    },
+    {
+      "epoch": 0.13407894736842105,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 3.111881782603852e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1352,
+      "loss/crossentropy": 2.1411483764648436,
+      "loss/hidden": 3.09375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2837516859173775,
+      "loss/reg": 0.0,
+      "step": 20380
+    },
+    {
+      "epoch": 0.13414473684210526,
+      "grad_norm": 7.5,
+      "grad_norm_var": 1.6672810872395833,
+      "learning_rate": 0.0001,
+      "loss": 3.139,
+      "loss/crossentropy": 2.224585199356079,
+      "loss/hidden": 2.9859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24173519760370255,
+      "loss/reg": 0.0,
+      "step": 20390
+    },
+    {
+      "epoch": 0.13421052631578947,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 1.8035634358723958,
+      "learning_rate": 0.0001,
+      "loss": 3.0216,
+      "loss/crossentropy": 2.153792452812195,
+      "loss/hidden": 2.6015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19605738371610643,
+      "loss/reg": 0.0,
+      "step": 20400
+    },
+    {
+      "epoch": 0.1342763157894737,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.18290913899739583,
+      "learning_rate": 0.0001,
+      "loss": 3.1134,
+      "loss/crossentropy": 2.3031864166259766,
+      "loss/hidden": 3.05,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26905038952827454,
+      "loss/reg": 0.0,
+      "step": 20410
+    },
+    {
+      "epoch": 0.1343421052631579,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.04761454264322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0931,
+      "loss/crossentropy": 2.250312161445618,
+      "loss/hidden": 2.9328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2365841895341873,
+      "loss/reg": 0.0,
+      "step": 20420
+    },
+    {
+      "epoch": 0.13440789473684212,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.06604410807291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0394,
+      "loss/crossentropy": 2.1318182408809663,
+      "loss/hidden": 3.0203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2652772217988968,
+      "loss/reg": 0.0,
+      "step": 20430
+    },
+    {
+      "epoch": 0.13447368421052633,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.054585774739583336,
+      "learning_rate": 0.0001,
+      "loss": 2.9721,
+      "loss/crossentropy": 2.237068510055542,
+      "loss/hidden": 2.65625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21526216119527816,
+      "loss/reg": 0.0,
+      "step": 20440
+    },
+    {
+      "epoch": 0.13453947368421051,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.05564676920572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0973,
+      "loss/crossentropy": 2.0180070281028746,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22148093655705453,
+      "loss/reg": 0.0,
+      "step": 20450
+    },
+    {
+      "epoch": 0.13460526315789473,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.058821614583333334,
+      "learning_rate": 0.0001,
+      "loss": 3.176,
+      "loss/crossentropy": 2.1725693702697755,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26163683980703356,
+      "loss/reg": 0.0,
+      "step": 20460
+    },
+    {
+      "epoch": 0.13467105263157894,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.06783625284830729,
+      "learning_rate": 0.0001,
+      "loss": 3.0549,
+      "loss/crossentropy": 1.9785831272602081,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21865816414356232,
+      "loss/reg": 0.0,
+      "step": 20470
+    },
+    {
+      "epoch": 0.13473684210526315,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.12560806274414063,
+      "learning_rate": 0.0001,
+      "loss": 3.1254,
+      "loss/crossentropy": 2.316719186306,
+      "loss/hidden": 2.690625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23244183510541916,
+      "loss/reg": 0.0,
+      "step": 20480
+    },
+    {
+      "epoch": 0.13480263157894737,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.1337554931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0912,
+      "loss/crossentropy": 2.2532522082328796,
+      "loss/hidden": 2.9703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2548024535179138,
+      "loss/reg": 0.0,
+      "step": 20490
+    },
+    {
+      "epoch": 0.13486842105263158,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.05217692057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0455,
+      "loss/crossentropy": 2.3155321717262267,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26803631633520125,
+      "loss/reg": 0.0,
+      "step": 20500
+    },
+    {
+      "epoch": 0.1349342105263158,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.4710896809895833,
+      "learning_rate": 0.0001,
+      "loss": 3.1847,
+      "loss/crossentropy": 2.4292261719703676,
+      "loss/hidden": 2.890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26074815839529036,
+      "loss/reg": 0.0,
+      "step": 20510
+    },
+    {
+      "epoch": 0.135,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.06902567545572917,
+      "learning_rate": 0.0001,
+      "loss": 3.1187,
+      "loss/crossentropy": 2.311373507976532,
+      "loss/hidden": 2.9453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2725979134440422,
+      "loss/reg": 0.0,
+      "step": 20520
+    },
+    {
+      "epoch": 0.13506578947368422,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.28644917805989584,
+      "learning_rate": 0.0001,
+      "loss": 3.0904,
+      "loss/crossentropy": 1.8008847087621689,
+      "loss/hidden": 2.8953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2606606900691986,
+      "loss/reg": 0.0,
+      "step": 20530
+    },
+    {
+      "epoch": 0.1351315789473684,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.06108296712239583,
+      "learning_rate": 0.0001,
+      "loss": 3.1547,
+      "loss/crossentropy": 2.0716704607009886,
+      "loss/hidden": 2.9484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21668420732021332,
+      "loss/reg": 0.0,
+      "step": 20540
+    },
+    {
+      "epoch": 0.13519736842105262,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.35935872395833335,
+      "learning_rate": 0.0001,
+      "loss": 3.1812,
+      "loss/crossentropy": 2.343623089790344,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25094970166683195,
+      "loss/reg": 0.0,
+      "step": 20550
+    },
+    {
+      "epoch": 0.13526315789473684,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.20611979166666666,
+      "learning_rate": 0.0001,
+      "loss": 3.1019,
+      "loss/crossentropy": 2.1298214733600616,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3111059933900833,
+      "loss/reg": 0.0,
+      "step": 20560
+    },
+    {
+      "epoch": 0.13532894736842105,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.07794596354166666,
+      "learning_rate": 0.0001,
+      "loss": 3.0833,
+      "loss/crossentropy": 2.1468781232833862,
+      "loss/hidden": 2.9421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.242007839679718,
+      "loss/reg": 0.0,
+      "step": 20570
+    },
+    {
+      "epoch": 0.13539473684210526,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.3861223856608073,
+      "learning_rate": 0.0001,
+      "loss": 3.056,
+      "loss/crossentropy": 2.0755307257175444,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21787920594215393,
+      "loss/reg": 0.0,
+      "step": 20580
+    },
+    {
+      "epoch": 0.13546052631578948,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.048713938395182295,
+      "learning_rate": 0.0001,
+      "loss": 3.0617,
+      "loss/crossentropy": 2.167508864402771,
+      "loss/hidden": 3.021875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2543604165315628,
+      "loss/reg": 0.0,
+      "step": 20590
+    },
+    {
+      "epoch": 0.1355263157894737,
+      "grad_norm": 4.40625,
+      "grad_norm_var": 0.31427408854166666,
+      "learning_rate": 0.0001,
+      "loss": 3.0881,
+      "loss/crossentropy": 2.1288257122039793,
+      "loss/hidden": 2.83125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2275502547621727,
+      "loss/reg": 0.0,
+      "step": 20600
+    },
+    {
+      "epoch": 0.1355921052631579,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.3212257385253906,
+      "learning_rate": 0.0001,
+      "loss": 3.0587,
+      "loss/crossentropy": 2.373918867111206,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25912273228168486,
+      "loss/reg": 0.0,
+      "step": 20610
+    },
+    {
+      "epoch": 0.13565789473684212,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.05368245442708333,
+      "learning_rate": 0.0001,
+      "loss": 3.0912,
+      "loss/crossentropy": 2.0399203658103944,
+      "loss/hidden": 2.884375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23557177633047105,
+      "loss/reg": 0.0,
+      "step": 20620
+    },
+    {
+      "epoch": 0.1357236842105263,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.11100031534830729,
+      "learning_rate": 0.0001,
+      "loss": 3.1051,
+      "loss/crossentropy": 2.5496551632881164,
+      "loss/hidden": 3.0109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2645132452249527,
+      "loss/reg": 0.0,
+      "step": 20630
+    },
+    {
+      "epoch": 0.13578947368421052,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.3220855712890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0665,
+      "loss/crossentropy": 2.226586413383484,
+      "loss/hidden": 2.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22172508686780928,
+      "loss/reg": 0.0,
+      "step": 20640
+    },
+    {
+      "epoch": 0.13585526315789473,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.2806925455729167,
+      "learning_rate": 0.0001,
+      "loss": 3.1914,
+      "loss/crossentropy": 2.2754149079322814,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27330992817878724,
+      "loss/reg": 0.0,
+      "step": 20650
+    },
+    {
+      "epoch": 0.13592105263157894,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.05778172810872396,
+      "learning_rate": 0.0001,
+      "loss": 3.1347,
+      "loss/crossentropy": 2.1621429324150085,
+      "loss/hidden": 2.8578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2567806988954544,
+      "loss/reg": 0.0,
+      "step": 20660
+    },
+    {
+      "epoch": 0.13598684210526316,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.2951637268066406,
+      "learning_rate": 0.0001,
+      "loss": 3.0649,
+      "loss/crossentropy": 2.342224645614624,
+      "loss/hidden": 2.7078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2490961804986,
+      "loss/reg": 0.0,
+      "step": 20670
+    },
+    {
+      "epoch": 0.13605263157894737,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.27790425618489584,
+      "learning_rate": 0.0001,
+      "loss": 3.1081,
+      "loss/crossentropy": 2.522631120681763,
+      "loss/hidden": 2.934375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28559967428445815,
+      "loss/reg": 0.0,
+      "step": 20680
+    },
+    {
+      "epoch": 0.13611842105263158,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.07203369140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0501,
+      "loss/crossentropy": 2.1872968673706055,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21910873502492906,
+      "loss/reg": 0.0,
+      "step": 20690
+    },
+    {
+      "epoch": 0.1361842105263158,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.06353251139322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0469,
+      "loss/crossentropy": 2.0313059210777284,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21065909266471863,
+      "loss/reg": 0.0,
+      "step": 20700
+    },
+    {
+      "epoch": 0.13625,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.12810872395833334,
+      "learning_rate": 0.0001,
+      "loss": 3.1641,
+      "loss/crossentropy": 1.9590769171714784,
+      "loss/hidden": 2.8453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2205381214618683,
+      "loss/reg": 0.0,
+      "step": 20710
+    },
+    {
+      "epoch": 0.13631578947368422,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.039526112874348956,
+      "learning_rate": 0.0001,
+      "loss": 3.0038,
+      "loss/crossentropy": 2.181920811533928,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20574837550520897,
+      "loss/reg": 0.0,
+      "step": 20720
+    },
+    {
+      "epoch": 0.1363815789473684,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.04170710245768229,
+      "learning_rate": 0.0001,
+      "loss": 3.0758,
+      "loss/crossentropy": 2.5629841327667235,
+      "loss/hidden": 2.7578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25942468345165254,
+      "loss/reg": 0.0,
+      "step": 20730
+    },
+    {
+      "epoch": 0.13644736842105262,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.04888916015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1239,
+      "loss/crossentropy": 2.5672937512397764,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2589055135846138,
+      "loss/reg": 0.0,
+      "step": 20740
+    },
+    {
+      "epoch": 0.13651315789473684,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.051558430989583334,
+      "learning_rate": 0.0001,
+      "loss": 3.1031,
+      "loss/crossentropy": 2.598326253890991,
+      "loss/hidden": 2.678125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2481001317501068,
+      "loss/reg": 0.0,
+      "step": 20750
+    },
+    {
+      "epoch": 0.13657894736842105,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.08925374348958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0563,
+      "loss/crossentropy": 2.403074288368225,
+      "loss/hidden": 2.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25444435328245163,
+      "loss/reg": 0.0,
+      "step": 20760
+    },
+    {
+      "epoch": 0.13664473684210526,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.33211161295572916,
+      "learning_rate": 0.0001,
+      "loss": 3.1043,
+      "loss/crossentropy": 2.161120080947876,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2631052315235138,
+      "loss/reg": 0.0,
+      "step": 20770
+    },
+    {
+      "epoch": 0.13671052631578948,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.2020416259765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0514,
+      "loss/crossentropy": 2.601893973350525,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23912052661180497,
+      "loss/reg": 0.0,
+      "step": 20780
+    },
+    {
+      "epoch": 0.1367763157894737,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.11573893229166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0923,
+      "loss/crossentropy": 2.3733190536499023,
+      "loss/hidden": 2.66875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21903542578220367,
+      "loss/reg": 0.0,
+      "step": 20790
+    },
+    {
+      "epoch": 0.1368421052631579,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.08520406087239583,
+      "learning_rate": 0.0001,
+      "loss": 3.1174,
+      "loss/crossentropy": 2.344668173789978,
+      "loss/hidden": 2.9328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2994943633675575,
+      "loss/reg": 0.0,
+      "step": 20800
+    },
+    {
+      "epoch": 0.13690789473684212,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 3.350255903434211e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2853,
+      "loss/crossentropy": 2.1936369478702544,
+      "loss/hidden": 2.9328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2379646047949791,
+      "loss/reg": 0.0,
+      "step": 20810
+    },
+    {
+      "epoch": 0.1369736842105263,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.025055948893229166,
+      "learning_rate": 0.0001,
+      "loss": 3.103,
+      "loss/crossentropy": 2.2369776248931883,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2727003887295723,
+      "loss/reg": 0.0,
+      "step": 20820
+    },
+    {
+      "epoch": 0.13703947368421052,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.033788045247395836,
+      "learning_rate": 0.0001,
+      "loss": 3.0878,
+      "loss/crossentropy": 2.2766017496585844,
+      "loss/hidden": 2.6875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21165435910224914,
+      "loss/reg": 0.0,
+      "step": 20830
+    },
+    {
+      "epoch": 0.13710526315789473,
+      "grad_norm": 5.84375,
+      "grad_norm_var": 0.75269775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1106,
+      "loss/crossentropy": 2.123941707611084,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2290189027786255,
+      "loss/reg": 0.0,
+      "step": 20840
+    },
+    {
+      "epoch": 0.13717105263157894,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.7671132405598958,
+      "learning_rate": 0.0001,
+      "loss": 3.0839,
+      "loss/crossentropy": 2.239221286773682,
+      "loss/hidden": 2.8390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2762044548988342,
+      "loss/reg": 0.0,
+      "step": 20850
+    },
+    {
+      "epoch": 0.13723684210526316,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.1312896728515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1207,
+      "loss/crossentropy": 2.4664002418518067,
+      "loss/hidden": 2.659375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2304532825946808,
+      "loss/reg": 0.0,
+      "step": 20860
+    },
+    {
+      "epoch": 0.13730263157894737,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.1725847880045573,
+      "learning_rate": 0.0001,
+      "loss": 3.1318,
+      "loss/crossentropy": 2.3205449104309084,
+      "loss/hidden": 2.715625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22169369757175444,
+      "loss/reg": 0.0,
+      "step": 20870
+    },
+    {
+      "epoch": 0.13736842105263158,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.046533203125,
+      "learning_rate": 0.0001,
+      "loss": 3.1764,
+      "loss/crossentropy": 2.2292275547981264,
+      "loss/hidden": 3.003125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26578798294067385,
+      "loss/reg": 0.0,
+      "step": 20880
+    },
+    {
+      "epoch": 0.1374342105263158,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.013053385416666667,
+      "learning_rate": 0.0001,
+      "loss": 3.0906,
+      "loss/crossentropy": 2.4730883955955507,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2600066691637039,
+      "loss/reg": 0.0,
+      "step": 20890
+    },
+    {
+      "epoch": 0.1375,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.03638916015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1146,
+      "loss/crossentropy": 2.389441525936127,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2204804763197899,
+      "loss/reg": 0.0,
+      "step": 20900
+    },
+    {
+      "epoch": 0.1375657894736842,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.05347391764322917,
+      "learning_rate": 0.0001,
+      "loss": 3.1005,
+      "loss/crossentropy": 2.182996892929077,
+      "loss/hidden": 2.9453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23823060542345048,
+      "loss/reg": 0.0,
+      "step": 20910
+    },
+    {
+      "epoch": 0.1376315789473684,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.0274810791015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1323,
+      "loss/crossentropy": 2.3811920285224915,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22427577376365662,
+      "loss/reg": 0.0,
+      "step": 20920
+    },
+    {
+      "epoch": 0.13769736842105262,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.0668609619140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1591,
+      "loss/crossentropy": 2.106644082069397,
+      "loss/hidden": 2.8890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27698966562747956,
+      "loss/reg": 0.0,
+      "step": 20930
+    },
+    {
+      "epoch": 0.13776315789473684,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.07274983723958334,
+      "learning_rate": 0.0001,
+      "loss": 3.165,
+      "loss/crossentropy": 2.3118181228637695,
+      "loss/hidden": 2.9078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30020255893468856,
+      "loss/reg": 0.0,
+      "step": 20940
+    },
+    {
+      "epoch": 0.13782894736842105,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.025804646809895835,
+      "learning_rate": 0.0001,
+      "loss": 3.1041,
+      "loss/crossentropy": 2.429260182380676,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2653680741786957,
+      "loss/reg": 0.0,
+      "step": 20950
+    },
+    {
+      "epoch": 0.13789473684210526,
+      "grad_norm": 3.25,
+      "grad_norm_var": 0.0960528055826823,
+      "learning_rate": 0.0001,
+      "loss": 3.0853,
+      "loss/crossentropy": 1.888709381222725,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21381159387528897,
+      "loss/reg": 0.0,
+      "step": 20960
+    },
+    {
+      "epoch": 0.13796052631578948,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.09798355102539062,
+      "learning_rate": 0.0001,
+      "loss": 3.1087,
+      "loss/crossentropy": 2.210513544082642,
+      "loss/hidden": 2.8390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24400804787874222,
+      "loss/reg": 0.0,
+      "step": 20970
+    },
+    {
+      "epoch": 0.1380263157894737,
+      "grad_norm": 2.0,
+      "grad_norm_var": 0.1358062744140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1309,
+      "loss/crossentropy": 2.3269863963127135,
+      "loss/hidden": 2.8640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25037053376436236,
+      "loss/reg": 0.0,
+      "step": 20980
+    },
+    {
+      "epoch": 0.1380921052631579,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.13828125,
+      "learning_rate": 0.0001,
+      "loss": 3.103,
+      "loss/crossentropy": 2.4014554262161254,
+      "loss/hidden": 2.703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22934938669204713,
+      "loss/reg": 0.0,
+      "step": 20990
+    },
+    {
+      "epoch": 0.13815789473684212,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.0584381103515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1088,
+      "loss/crossentropy": 2.208330225944519,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23984504938125611,
+      "loss/reg": 0.0,
+      "step": 21000
+    },
+    {
+      "epoch": 0.1382236842105263,
+      "grad_norm": 1.9296875,
+      "grad_norm_var": 0.1774614969889323,
+      "learning_rate": 0.0001,
+      "loss": 3.0588,
+      "loss/crossentropy": 2.496232438087463,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29541969746351243,
+      "loss/reg": 0.0,
+      "step": 21010
+    },
+    {
+      "epoch": 0.13828947368421052,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.4096514383951823,
+      "learning_rate": 0.0001,
+      "loss": 3.1337,
+      "loss/crossentropy": 2.167738914489746,
+      "loss/hidden": 2.9234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23880672752857207,
+      "loss/reg": 0.0,
+      "step": 21020
+    },
+    {
+      "epoch": 0.13835526315789473,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.06155192057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1325,
+      "loss/crossentropy": 2.1132714807987214,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.227035228908062,
+      "loss/reg": 0.0,
+      "step": 21030
+    },
+    {
+      "epoch": 0.13842105263157894,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.09390360514322917,
+      "learning_rate": 0.0001,
+      "loss": 3.1269,
+      "loss/crossentropy": 2.467393732070923,
+      "loss/hidden": 2.91875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24615048468112946,
+      "loss/reg": 0.0,
+      "step": 21040
+    },
+    {
+      "epoch": 0.13848684210526316,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.035090128580729164,
+      "learning_rate": 0.0001,
+      "loss": 3.1166,
+      "loss/crossentropy": 2.2899758577346803,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25930332988500593,
+      "loss/reg": 0.0,
+      "step": 21050
+    },
+    {
+      "epoch": 0.13855263157894737,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.08716812133789062,
+      "learning_rate": 0.0001,
+      "loss": 3.074,
+      "loss/crossentropy": 2.3258708715438843,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23002547025680542,
+      "loss/reg": 0.0,
+      "step": 21060
+    },
+    {
+      "epoch": 0.13861842105263159,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.10960667928059896,
+      "learning_rate": 0.0001,
+      "loss": 3.1011,
+      "loss/crossentropy": 2.368458116054535,
+      "loss/hidden": 2.85625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26667492985725405,
+      "loss/reg": 0.0,
+      "step": 21070
+    },
+    {
+      "epoch": 0.1386842105263158,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.053807576497395836,
+      "learning_rate": 0.0001,
+      "loss": 3.0992,
+      "loss/crossentropy": 2.268073225021362,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23710739463567734,
+      "loss/reg": 0.0,
+      "step": 21080
+    },
+    {
+      "epoch": 0.13875,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.7524088541666667,
+      "learning_rate": 0.0001,
+      "loss": 3.123,
+      "loss/crossentropy": 2.116519570350647,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2593864217400551,
+      "loss/reg": 0.0,
+      "step": 21090
+    },
+    {
+      "epoch": 0.1388157894736842,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.10280329386393229,
+      "learning_rate": 0.0001,
+      "loss": 3.0788,
+      "loss/crossentropy": 2.49674973487854,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2637839734554291,
+      "loss/reg": 0.0,
+      "step": 21100
+    },
+    {
+      "epoch": 0.1388815789473684,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.06420059204101562,
+      "learning_rate": 0.0001,
+      "loss": 3.0901,
+      "loss/crossentropy": 2.142958414554596,
+      "loss/hidden": 2.9453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28233732134103773,
+      "loss/reg": 0.0,
+      "step": 21110
+    },
+    {
+      "epoch": 0.13894736842105262,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.09973551432291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1008,
+      "loss/crossentropy": 2.300196385383606,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24371459782123567,
+      "loss/reg": 0.0,
+      "step": 21120
+    },
+    {
+      "epoch": 0.13901315789473684,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.07563247680664062,
+      "learning_rate": 0.0001,
+      "loss": 3.0722,
+      "loss/crossentropy": 2.424953353404999,
+      "loss/hidden": 3.08125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28959580361843107,
+      "loss/reg": 0.0,
+      "step": 21130
+    },
+    {
+      "epoch": 0.13907894736842105,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.05977554321289062,
+      "learning_rate": 0.0001,
+      "loss": 3.1188,
+      "loss/crossentropy": 2.3048367381095884,
+      "loss/hidden": 2.9796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26107140332460405,
+      "loss/reg": 0.0,
+      "step": 21140
+    },
+    {
+      "epoch": 0.13914473684210527,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.17164713541666668,
+      "learning_rate": 0.0001,
+      "loss": 3.1765,
+      "loss/crossentropy": 2.376292657852173,
+      "loss/hidden": 2.753125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24798475652933122,
+      "loss/reg": 0.0,
+      "step": 21150
+    },
+    {
+      "epoch": 0.13921052631578948,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.0533355712890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0556,
+      "loss/crossentropy": 2.620091509819031,
+      "loss/hidden": 2.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2451164111495018,
+      "loss/reg": 0.0,
+      "step": 21160
+    },
+    {
+      "epoch": 0.1392763157894737,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.07333882649739583,
+      "learning_rate": 0.0001,
+      "loss": 3.1244,
+      "loss/crossentropy": 2.2057873249053954,
+      "loss/hidden": 2.9109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26459684371948244,
+      "loss/reg": 0.0,
+      "step": 21170
+    },
+    {
+      "epoch": 0.1393421052631579,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.024214680989583334,
+      "learning_rate": 0.0001,
+      "loss": 3.0668,
+      "loss/crossentropy": 2.087160420417786,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2084551602602005,
+      "loss/reg": 0.0,
+      "step": 21180
+    },
+    {
+      "epoch": 0.1394078947368421,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.10177408854166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0369,
+      "loss/crossentropy": 2.4722538352012635,
+      "loss/hidden": 2.959375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28627206683158873,
+      "loss/reg": 0.0,
+      "step": 21190
+    },
+    {
+      "epoch": 0.1394736842105263,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 3.921613566080729,
+      "learning_rate": 0.0001,
+      "loss": 3.2147,
+      "loss/crossentropy": 2.103289079666138,
+      "loss/hidden": 2.921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2449532501399517,
+      "loss/reg": 0.0,
+      "step": 21200
+    },
+    {
+      "epoch": 0.13953947368421052,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 3.8165679931640626,
+      "learning_rate": 0.0001,
+      "loss": 3.123,
+      "loss/crossentropy": 2.377441930770874,
+      "loss/hidden": 2.8875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27957661896944047,
+      "loss/reg": 0.0,
+      "step": 21210
+    },
+    {
+      "epoch": 0.13960526315789473,
+      "grad_norm": 3.34375,
+      "grad_norm_var": 0.18837788899739583,
+      "learning_rate": 0.0001,
+      "loss": 3.1171,
+      "loss/crossentropy": 2.218759763240814,
+      "loss/hidden": 2.7578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2158270835876465,
+      "loss/reg": 0.0,
+      "step": 21220
+    },
+    {
+      "epoch": 0.13967105263157895,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 14.1974609375,
+      "learning_rate": 0.0001,
+      "loss": 3.1327,
+      "loss/crossentropy": 2.4489439368247985,
+      "loss/hidden": 2.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23917225003242493,
+      "loss/reg": 0.0,
+      "step": 21230
+    },
+    {
+      "epoch": 0.13973684210526316,
+      "grad_norm": 3.140625,
+      "grad_norm_var": 14.437064361572265,
+      "learning_rate": 0.0001,
+      "loss": 3.0414,
+      "loss/crossentropy": 2.2649194478988646,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21438237726688386,
+      "loss/reg": 0.0,
+      "step": 21240
+    },
+    {
+      "epoch": 0.13980263157894737,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 2.151969401041667,
+      "learning_rate": 0.0001,
+      "loss": 3.1748,
+      "loss/crossentropy": 2.350678837299347,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2586789205670357,
+      "loss/reg": 0.0,
+      "step": 21250
+    },
+    {
+      "epoch": 0.1398684210526316,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.07167867024739584,
+      "learning_rate": 0.0001,
+      "loss": 3.1436,
+      "loss/crossentropy": 2.383778703212738,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30858664512634276,
+      "loss/reg": 0.0,
+      "step": 21260
+    },
+    {
+      "epoch": 0.1399342105263158,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.04163004557291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0829,
+      "loss/crossentropy": 2.4344274759292603,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24489458352327348,
+      "loss/reg": 0.0,
+      "step": 21270
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.04241129557291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0751,
+      "loss/crossentropy": 2.59662504196167,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23801603466272353,
+      "loss/reg": 0.0,
+      "step": 21280
+    },
+    {
+      "epoch": 0.1400657894736842,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 1.7194010416666667,
+      "learning_rate": 0.0001,
+      "loss": 3.0867,
+      "loss/crossentropy": 2.466573119163513,
+      "loss/hidden": 2.9375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29062790870666505,
+      "loss/reg": 0.0,
+      "step": 21290
+    },
+    {
+      "epoch": 0.1401315789473684,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.19259440104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.2042,
+      "loss/crossentropy": 2.5092820644378664,
+      "loss/hidden": 3.0046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3163674846291542,
+      "loss/reg": 0.0,
+      "step": 21300
+    },
+    {
+      "epoch": 0.14019736842105263,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.09862874348958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0485,
+      "loss/crossentropy": 2.1548155784606933,
+      "loss/hidden": 2.64375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20280121639370918,
+      "loss/reg": 0.0,
+      "step": 21310
+    },
+    {
+      "epoch": 0.14026315789473684,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.25152587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0701,
+      "loss/crossentropy": 2.1562862396240234,
+      "loss/hidden": 2.740625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1931696727871895,
+      "loss/reg": 0.0,
+      "step": 21320
+    },
+    {
+      "epoch": 0.14032894736842105,
+      "grad_norm": 7.6875,
+      "grad_norm_var": 1.9003896077473958,
+      "learning_rate": 0.0001,
+      "loss": 3.1242,
+      "loss/crossentropy": 2.4868319749832155,
+      "loss/hidden": 3.2,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.263777095079422,
+      "loss/reg": 0.0,
+      "step": 21330
+    },
+    {
+      "epoch": 0.14039473684210527,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 2.187555948893229,
+      "learning_rate": 0.0001,
+      "loss": 3.0713,
+      "loss/crossentropy": 2.4403932213783266,
+      "loss/hidden": 2.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.239518903195858,
+      "loss/reg": 0.0,
+      "step": 21340
+    },
+    {
+      "epoch": 0.14046052631578948,
+      "grad_norm": 2.9375,
+      "grad_norm_var": 0.92603759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.161,
+      "loss/crossentropy": 2.637459659576416,
+      "loss/hidden": 2.846875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2770415723323822,
+      "loss/reg": 0.0,
+      "step": 21350
+    },
+    {
+      "epoch": 0.1405263157894737,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.6560373942057292,
+      "learning_rate": 0.0001,
+      "loss": 3.1023,
+      "loss/crossentropy": 2.4345417737960817,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23416574895381928,
+      "loss/reg": 0.0,
+      "step": 21360
+    },
+    {
+      "epoch": 0.1405921052631579,
+      "grad_norm": 4.625,
+      "grad_norm_var": 0.39641927083333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0819,
+      "loss/crossentropy": 2.3080653309822083,
+      "loss/hidden": 2.6703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.203302001953125,
+      "loss/reg": 0.0,
+      "step": 21370
+    },
+    {
+      "epoch": 0.1406578947368421,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.3663736979166667,
+      "learning_rate": 0.0001,
+      "loss": 3.005,
+      "loss/crossentropy": 2.312281048297882,
+      "loss/hidden": 2.75625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22557048946619035,
+      "loss/reg": 0.0,
+      "step": 21380
+    },
+    {
+      "epoch": 0.1407236842105263,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.0878326416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0862,
+      "loss/crossentropy": 2.725596809387207,
+      "loss/hidden": 2.83125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26989305168390276,
+      "loss/reg": 0.0,
+      "step": 21390
+    },
+    {
+      "epoch": 0.14078947368421052,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.7812662760416667,
+      "learning_rate": 0.0001,
+      "loss": 3.1293,
+      "loss/crossentropy": 2.4127392411231994,
+      "loss/hidden": 2.9703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28430653512477877,
+      "loss/reg": 0.0,
+      "step": 21400
+    },
+    {
+      "epoch": 0.14085526315789473,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.76865234375,
+      "learning_rate": 0.0001,
+      "loss": 3.1054,
+      "loss/crossentropy": 2.3319249004125595,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22504624500870704,
+      "loss/reg": 0.0,
+      "step": 21410
+    },
+    {
+      "epoch": 0.14092105263157895,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.09716389973958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0617,
+      "loss/crossentropy": 2.6645477771759034,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24796359091997147,
+      "loss/reg": 0.0,
+      "step": 21420
+    },
+    {
+      "epoch": 0.14098684210526316,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 4.503599619195098e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2735,
+      "loss/crossentropy": 2.1648690342903136,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23503894209861756,
+      "loss/reg": 0.0,
+      "step": 21430
+    },
+    {
+      "epoch": 0.14105263157894737,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.07038472493489584,
+      "learning_rate": 0.0001,
+      "loss": 3.0659,
+      "loss/crossentropy": 2.477909338474274,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2327214926481247,
+      "loss/reg": 0.0,
+      "step": 21440
+    },
+    {
+      "epoch": 0.1411184210526316,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.03515625,
+      "learning_rate": 0.0001,
+      "loss": 3.057,
+      "loss/crossentropy": 2.354469347000122,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23355985432863235,
+      "loss/reg": 0.0,
+      "step": 21450
+    },
+    {
+      "epoch": 0.1411842105263158,
+      "grad_norm": 3.34375,
+      "grad_norm_var": 0.18144124348958332,
+      "learning_rate": 0.0001,
+      "loss": 3.1157,
+      "loss/crossentropy": 2.2400832891464235,
+      "loss/hidden": 2.8390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24985998570919038,
+      "loss/reg": 0.0,
+      "step": 21460
+    },
+    {
+      "epoch": 0.14125,
+      "grad_norm": 5.65625,
+      "grad_norm_var": 0.7304921468098958,
+      "learning_rate": 0.0001,
+      "loss": 3.0675,
+      "loss/crossentropy": 2.324231135845184,
+      "loss/hidden": 2.6875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22793666571378707,
+      "loss/reg": 0.0,
+      "step": 21470
+    },
+    {
+      "epoch": 0.1413157894736842,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.7799641927083333,
+      "learning_rate": 0.0001,
+      "loss": 3.1565,
+      "loss/crossentropy": 2.4094609022140503,
+      "loss/hidden": 2.809375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21761243641376496,
+      "loss/reg": 0.0,
+      "step": 21480
+    },
+    {
+      "epoch": 0.1413815789473684,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.1587310791015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0847,
+      "loss/crossentropy": 1.9431841492652893,
+      "loss/hidden": 2.846875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24099752232432364,
+      "loss/reg": 0.0,
+      "step": 21490
+    },
+    {
+      "epoch": 0.14144736842105263,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.0576324462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0518,
+      "loss/crossentropy": 2.5469526767730715,
+      "loss/hidden": 2.7234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24212357103824617,
+      "loss/reg": 0.0,
+      "step": 21500
+    },
+    {
+      "epoch": 0.14151315789473684,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.07248942057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1081,
+      "loss/crossentropy": 2.303622233867645,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23643973991274833,
+      "loss/reg": 0.0,
+      "step": 21510
+    },
+    {
+      "epoch": 0.14157894736842105,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.16367085774739584,
+      "learning_rate": 0.0001,
+      "loss": 3.0839,
+      "loss/crossentropy": 2.178890359401703,
+      "loss/hidden": 2.8265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23347833156585693,
+      "loss/reg": 0.0,
+      "step": 21520
+    },
+    {
+      "epoch": 0.14164473684210527,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.0575836181640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1122,
+      "loss/crossentropy": 2.4304057359695435,
+      "loss/hidden": 2.725,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2439558282494545,
+      "loss/reg": 0.0,
+      "step": 21530
+    },
+    {
+      "epoch": 0.14171052631578948,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.1728668212890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1416,
+      "loss/crossentropy": 2.111624151468277,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2380165532231331,
+      "loss/reg": 0.0,
+      "step": 21540
+    },
+    {
+      "epoch": 0.1417763157894737,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.17907613118489582,
+      "learning_rate": 0.0001,
+      "loss": 3.0817,
+      "loss/crossentropy": 2.2606616258621215,
+      "loss/hidden": 3.0546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24319592714309693,
+      "loss/reg": 0.0,
+      "step": 21550
+    },
+    {
+      "epoch": 0.1418421052631579,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.065234375,
+      "learning_rate": 0.0001,
+      "loss": 3.0592,
+      "loss/crossentropy": 2.2833418786525725,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24779771864414216,
+      "loss/reg": 0.0,
+      "step": 21560
+    },
+    {
+      "epoch": 0.1419078947368421,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.19248758951822917,
+      "learning_rate": 0.0001,
+      "loss": 3.1323,
+      "loss/crossentropy": 2.2865166664123535,
+      "loss/hidden": 2.809375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2382546842098236,
+      "loss/reg": 0.0,
+      "step": 21570
+    },
+    {
+      "epoch": 0.1419736842105263,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.19429931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0883,
+      "loss/crossentropy": 2.37155544757843,
+      "loss/hidden": 3.0421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28304021507501603,
+      "loss/reg": 0.0,
+      "step": 21580
+    },
+    {
+      "epoch": 0.14203947368421052,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.06051432291666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1666,
+      "loss/crossentropy": 2.0952234268188477,
+      "loss/hidden": 2.8390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28353812396526334,
+      "loss/reg": 0.0,
+      "step": 21590
+    },
+    {
+      "epoch": 0.14210526315789473,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.0432769775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1277,
+      "loss/crossentropy": 2.6342113494873045,
+      "loss/hidden": 3.0078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27442471832036974,
+      "loss/reg": 0.0,
+      "step": 21600
+    },
+    {
+      "epoch": 0.14217105263157895,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.052018229166666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0889,
+      "loss/crossentropy": 2.3795015037059786,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24964673370122908,
+      "loss/reg": 0.0,
+      "step": 21610
+    },
+    {
+      "epoch": 0.14223684210526316,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.15912984212239584,
+      "learning_rate": 0.0001,
+      "loss": 3.1592,
+      "loss/crossentropy": 2.435245943069458,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.36159887462854384,
+      "loss/reg": 0.0,
+      "step": 21620
+    },
+    {
+      "epoch": 0.14230263157894738,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.19715067545572917,
+      "learning_rate": 0.0001,
+      "loss": 2.9695,
+      "loss/crossentropy": 2.3731093287467955,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24517869502305983,
+      "loss/reg": 0.0,
+      "step": 21630
+    },
+    {
+      "epoch": 0.1423684210526316,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.029588826497395835,
+      "learning_rate": 0.0001,
+      "loss": 3.09,
+      "loss/crossentropy": 2.1785697817802427,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22660856544971467,
+      "loss/reg": 0.0,
+      "step": 21640
+    },
+    {
+      "epoch": 0.1424342105263158,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.0379302978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1062,
+      "loss/crossentropy": 2.3706693768501284,
+      "loss/hidden": 2.9953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27286539524793624,
+      "loss/reg": 0.0,
+      "step": 21650
+    },
+    {
+      "epoch": 0.1425,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.05085347493489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1297,
+      "loss/crossentropy": 2.173750901222229,
+      "loss/hidden": 2.9984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2637738898396492,
+      "loss/reg": 0.0,
+      "step": 21660
+    },
+    {
+      "epoch": 0.1425657894736842,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.10258153279622396,
+      "learning_rate": 0.0001,
+      "loss": 3.0637,
+      "loss/crossentropy": 2.133163595199585,
+      "loss/hidden": 2.6578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19840268045663834,
+      "loss/reg": 0.0,
+      "step": 21670
+    },
+    {
+      "epoch": 0.14263157894736841,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.062646484375,
+      "learning_rate": 0.0001,
+      "loss": 3.0662,
+      "loss/crossentropy": 2.4523648262023925,
+      "loss/hidden": 2.684375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23104819357395173,
+      "loss/reg": 0.0,
+      "step": 21680
+    },
+    {
+      "epoch": 0.14269736842105263,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.2750935872395833,
+      "learning_rate": 0.0001,
+      "loss": 3.1536,
+      "loss/crossentropy": 2.250650954246521,
+      "loss/hidden": 2.86875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24096653312444688,
+      "loss/reg": 0.0,
+      "step": 21690
+    },
+    {
+      "epoch": 0.14276315789473684,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.26304931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0649,
+      "loss/crossentropy": 2.30162969827652,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24212537854909896,
+      "loss/reg": 0.0,
+      "step": 21700
+    },
+    {
+      "epoch": 0.14282894736842106,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.027521769205729168,
+      "learning_rate": 0.0001,
+      "loss": 3.1097,
+      "loss/crossentropy": 2.1112605214118956,
+      "loss/hidden": 3.04375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26635565906763076,
+      "loss/reg": 0.0,
+      "step": 21710
+    },
+    {
+      "epoch": 0.14289473684210527,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.0798492431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0973,
+      "loss/crossentropy": 2.3435486197471618,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2827204465866089,
+      "loss/reg": 0.0,
+      "step": 21720
+    },
+    {
+      "epoch": 0.14296052631578948,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.114111328125,
+      "learning_rate": 0.0001,
+      "loss": 3.1039,
+      "loss/crossentropy": 2.4754523396492005,
+      "loss/hidden": 2.9296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29244270324707033,
+      "loss/reg": 0.0,
+      "step": 21730
+    },
+    {
+      "epoch": 0.1430263157894737,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.18517252604166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1588,
+      "loss/crossentropy": 2.2757002115249634,
+      "loss/hidden": 2.9109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27034522593021393,
+      "loss/reg": 0.0,
+      "step": 21740
+    },
+    {
+      "epoch": 0.14309210526315788,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.19296773274739584,
+      "learning_rate": 0.0001,
+      "loss": 3.112,
+      "loss/crossentropy": 2.52050644159317,
+      "loss/hidden": 2.9640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2818415179848671,
+      "loss/reg": 0.0,
+      "step": 21750
+    },
+    {
+      "epoch": 0.1431578947368421,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.1665679931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.2325,
+      "loss/crossentropy": 2.3630916833877564,
+      "loss/hidden": 2.8203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2456425666809082,
+      "loss/reg": 0.0,
+      "step": 21760
+    },
+    {
+      "epoch": 0.1432236842105263,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 0.21158854166666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0907,
+      "loss/crossentropy": 2.1248918056488035,
+      "loss/hidden": 2.8640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22751243263483048,
+      "loss/reg": 0.0,
+      "step": 21770
+    },
+    {
+      "epoch": 0.14328947368421052,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.06939697265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0952,
+      "loss/crossentropy": 2.273802196979523,
+      "loss/hidden": 2.98125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2570581123232841,
+      "loss/reg": 0.0,
+      "step": 21780
+    },
+    {
+      "epoch": 0.14335526315789474,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.025763956705729167,
+      "learning_rate": 0.0001,
+      "loss": 3.0584,
+      "loss/crossentropy": 2.096750485897064,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21872427612543105,
+      "loss/reg": 0.0,
+      "step": 21790
+    },
+    {
+      "epoch": 0.14342105263157895,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.023412068684895832,
+      "learning_rate": 0.0001,
+      "loss": 3.0604,
+      "loss/crossentropy": 2.228634536266327,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2626715019345284,
+      "loss/reg": 0.0,
+      "step": 21800
+    },
+    {
+      "epoch": 0.14348684210526316,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.28734512329101564,
+      "learning_rate": 0.0001,
+      "loss": 3.1324,
+      "loss/crossentropy": 2.475996434688568,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23401851058006287,
+      "loss/reg": 0.0,
+      "step": 21810
+    },
+    {
+      "epoch": 0.14355263157894738,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.39009984334309894,
+      "learning_rate": 0.0001,
+      "loss": 3.1624,
+      "loss/crossentropy": 2.285959267616272,
+      "loss/hidden": 3.0015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24858510196208955,
+      "loss/reg": 0.0,
+      "step": 21820
+    },
+    {
+      "epoch": 0.1436184210526316,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.18899739583333333,
+      "learning_rate": 0.0001,
+      "loss": 3.0702,
+      "loss/crossentropy": 2.463927984237671,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2442393772304058,
+      "loss/reg": 0.0,
+      "step": 21830
+    },
+    {
+      "epoch": 0.1436842105263158,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.14226786295572916,
+      "learning_rate": 0.0001,
+      "loss": 3.0965,
+      "loss/crossentropy": 2.277498161792755,
+      "loss/hidden": 2.7578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2153932645916939,
+      "loss/reg": 0.0,
+      "step": 21840
+    },
+    {
+      "epoch": 0.14375,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 1.7546953837076822,
+      "learning_rate": 0.0001,
+      "loss": 3.108,
+      "loss/crossentropy": 2.3049885034561157,
+      "loss/hidden": 2.94375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26804344207048414,
+      "loss/reg": 0.0,
+      "step": 21850
+    },
+    {
+      "epoch": 0.1438157894736842,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 4.087398020426432,
+      "learning_rate": 0.0001,
+      "loss": 3.1586,
+      "loss/crossentropy": 2.239436650276184,
+      "loss/hidden": 3.0796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27900682389736176,
+      "loss/reg": 0.0,
+      "step": 21860
+    },
+    {
+      "epoch": 0.14388157894736842,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.09602762858072916,
+      "learning_rate": 0.0001,
+      "loss": 3.1657,
+      "loss/crossentropy": 2.385130798816681,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28002837896347044,
+      "loss/reg": 0.0,
+      "step": 21870
+    },
+    {
+      "epoch": 0.14394736842105263,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.12711588541666666,
+      "learning_rate": 0.0001,
+      "loss": 3.1437,
+      "loss/crossentropy": 2.426273798942566,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3090307667851448,
+      "loss/reg": 0.0,
+      "step": 21880
+    },
+    {
+      "epoch": 0.14401315789473684,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.09078776041666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1105,
+      "loss/crossentropy": 2.148479038476944,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22178827822208405,
+      "loss/reg": 0.0,
+      "step": 21890
+    },
+    {
+      "epoch": 0.14407894736842106,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.0625,
+      "learning_rate": 0.0001,
+      "loss": 3.1009,
+      "loss/crossentropy": 2.2989478588104246,
+      "loss/hidden": 2.8140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21844121366739272,
+      "loss/reg": 0.0,
+      "step": 21900
+    },
+    {
+      "epoch": 0.14414473684210527,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.10066630045572916,
+      "learning_rate": 0.0001,
+      "loss": 3.1368,
+      "loss/crossentropy": 2.459014880657196,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2873328909277916,
+      "loss/reg": 0.0,
+      "step": 21910
+    },
+    {
+      "epoch": 0.14421052631578948,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.15559794108072916,
+      "learning_rate": 0.0001,
+      "loss": 3.048,
+      "loss/crossentropy": 2.0745410680770875,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24865385442972182,
+      "loss/reg": 0.0,
+      "step": 21920
+    },
+    {
+      "epoch": 0.1442763157894737,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.1090240478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.2472,
+      "loss/crossentropy": 2.216059982776642,
+      "loss/hidden": 2.946875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25595484375953675,
+      "loss/reg": 0.0,
+      "step": 21930
+    },
+    {
+      "epoch": 0.14434210526315788,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.06485773722330729,
+      "learning_rate": 0.0001,
+      "loss": 3.0388,
+      "loss/crossentropy": 2.3124427676200865,
+      "loss/hidden": 2.678125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2212652564048767,
+      "loss/reg": 0.0,
+      "step": 21940
+    },
+    {
+      "epoch": 0.1444078947368421,
+      "grad_norm": 2.953125,
+      "grad_norm_var": 0.4722246805826823,
+      "learning_rate": 0.0001,
+      "loss": 3.2024,
+      "loss/crossentropy": 2.1076952695846556,
+      "loss/hidden": 2.946875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2454703077673912,
+      "loss/reg": 0.0,
+      "step": 21950
+    },
+    {
+      "epoch": 0.1444736842105263,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.6233306884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1576,
+      "loss/crossentropy": 2.3738736391067503,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22644958645105362,
+      "loss/reg": 0.0,
+      "step": 21960
+    },
+    {
+      "epoch": 0.14453947368421052,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.48958231608072916,
+      "learning_rate": 0.0001,
+      "loss": 3.1195,
+      "loss/crossentropy": 2.3681930541992187,
+      "loss/hidden": 2.828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23997818529605866,
+      "loss/reg": 0.0,
+      "step": 21970
+    },
+    {
+      "epoch": 0.14460526315789474,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.06328125,
+      "learning_rate": 0.0001,
+      "loss": 3.086,
+      "loss/crossentropy": 2.202842915058136,
+      "loss/hidden": 2.8578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22619501650333404,
+      "loss/reg": 0.0,
+      "step": 21980
+    },
+    {
+      "epoch": 0.14467105263157895,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.08901265462239584,
+      "learning_rate": 0.0001,
+      "loss": 3.0968,
+      "loss/crossentropy": 2.498619997501373,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2592793509364128,
+      "loss/reg": 0.0,
+      "step": 21990
+    },
+    {
+      "epoch": 0.14473684210526316,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.07700907389322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0646,
+      "loss/crossentropy": 2.5274386525154116,
+      "loss/hidden": 2.7109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2393401548266411,
+      "loss/reg": 0.0,
+      "step": 22000
+    },
+    {
+      "epoch": 0.14480263157894738,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.03200581868489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1375,
+      "loss/crossentropy": 2.331031596660614,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22175868153572081,
+      "loss/reg": 0.0,
+      "step": 22010
+    },
+    {
+      "epoch": 0.1448684210526316,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.031473541259765626,
+      "learning_rate": 0.0001,
+      "loss": 3.1187,
+      "loss/crossentropy": 2.1961219191551207,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23576426208019258,
+      "loss/reg": 0.0,
+      "step": 22020
+    },
+    {
+      "epoch": 0.14493421052631578,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.04307835896809896,
+      "learning_rate": 0.0001,
+      "loss": 3.0722,
+      "loss/crossentropy": 2.4104359984397887,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25036042332649233,
+      "loss/reg": 0.0,
+      "step": 22030
+    },
+    {
+      "epoch": 0.145,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.0566558837890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0535,
+      "loss/crossentropy": 2.293129473924637,
+      "loss/hidden": 2.909375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3266716688871384,
+      "loss/reg": 0.0,
+      "step": 22040
+    },
+    {
+      "epoch": 0.1450657894736842,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.1048828125,
+      "learning_rate": 0.0001,
+      "loss": 3.151,
+      "loss/crossentropy": 2.1948832154273985,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2283545032143593,
+      "loss/reg": 0.0,
+      "step": 22050
+    },
+    {
+      "epoch": 0.14513157894736842,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.85728759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1379,
+      "loss/crossentropy": 2.4909852266311647,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26901226192712785,
+      "loss/reg": 0.0,
+      "step": 22060
+    },
+    {
+      "epoch": 0.14519736842105263,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.8632476806640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0345,
+      "loss/crossentropy": 2.4894071340560915,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2518943950533867,
+      "loss/reg": 0.0,
+      "step": 22070
+    },
+    {
+      "epoch": 0.14526315789473684,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.035553995768229166,
+      "learning_rate": 0.0001,
+      "loss": 3.0777,
+      "loss/crossentropy": 2.250881004333496,
+      "loss/hidden": 2.9375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.254881876707077,
+      "loss/reg": 0.0,
+      "step": 22080
+    },
+    {
+      "epoch": 0.14532894736842106,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.6524251302083334,
+      "learning_rate": 0.0001,
+      "loss": 3.0046,
+      "loss/crossentropy": 2.4635193943977356,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28406478762626647,
+      "loss/reg": 0.0,
+      "step": 22090
+    },
+    {
+      "epoch": 0.14539473684210527,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.2197906494140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0607,
+      "loss/crossentropy": 2.176581871509552,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2112376168370247,
+      "loss/reg": 0.0,
+      "step": 22100
+    },
+    {
+      "epoch": 0.14546052631578948,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.0716461181640625,
+      "learning_rate": 0.0001,
+      "loss": 3.09,
+      "loss/crossentropy": 2.418153405189514,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24471160471439363,
+      "loss/reg": 0.0,
+      "step": 22110
+    },
+    {
+      "epoch": 0.1455263157894737,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.024909464518229167,
+      "learning_rate": 0.0001,
+      "loss": 3.0661,
+      "loss/crossentropy": 2.34368360042572,
+      "loss/hidden": 2.9484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26933753192424775,
+      "loss/reg": 0.0,
+      "step": 22120
+    },
+    {
+      "epoch": 0.14559210526315788,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.0403228759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0961,
+      "loss/crossentropy": 2.2858925461769104,
+      "loss/hidden": 2.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23935395181179048,
+      "loss/reg": 0.0,
+      "step": 22130
+    },
+    {
+      "epoch": 0.1456578947368421,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.06553446451822917,
+      "learning_rate": 0.0001,
+      "loss": 3.1287,
+      "loss/crossentropy": 2.476478910446167,
+      "loss/hidden": 2.821875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2711992233991623,
+      "loss/reg": 0.0,
+      "step": 22140
+    },
+    {
+      "epoch": 0.1457236842105263,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.14517822265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0753,
+      "loss/crossentropy": 2.095052421092987,
+      "loss/hidden": 2.95625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2310307502746582,
+      "loss/reg": 0.0,
+      "step": 22150
+    },
+    {
+      "epoch": 0.14578947368421052,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.13459243774414062,
+      "learning_rate": 0.0001,
+      "loss": 3.0609,
+      "loss/crossentropy": 2.252713418006897,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2497629001736641,
+      "loss/reg": 0.0,
+      "step": 22160
+    },
+    {
+      "epoch": 0.14585526315789474,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.06102701822916667,
+      "learning_rate": 0.0001,
+      "loss": 3.092,
+      "loss/crossentropy": 2.2792391180992126,
+      "loss/hidden": 2.9890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2850720778107643,
+      "loss/reg": 0.0,
+      "step": 22170
+    },
+    {
+      "epoch": 0.14592105263157895,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.2100738525390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1172,
+      "loss/crossentropy": 2.116334557533264,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22928555756807328,
+      "loss/reg": 0.0,
+      "step": 22180
+    },
+    {
+      "epoch": 0.14598684210526316,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.12631734212239584,
+      "learning_rate": 0.0001,
+      "loss": 3.103,
+      "loss/crossentropy": 2.2788902163505553,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24738141447305678,
+      "loss/reg": 0.0,
+      "step": 22190
+    },
+    {
+      "epoch": 0.14605263157894738,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.05478108723958333,
+      "learning_rate": 0.0001,
+      "loss": 2.9932,
+      "loss/crossentropy": 2.3781439542770384,
+      "loss/hidden": 2.9796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2298112317919731,
+      "loss/reg": 0.0,
+      "step": 22200
+    },
+    {
+      "epoch": 0.1461184210526316,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 1.2361399332682292,
+      "learning_rate": 0.0001,
+      "loss": 3.1279,
+      "loss/crossentropy": 2.367407274246216,
+      "loss/hidden": 3.0234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25443960130214693,
+      "loss/reg": 0.0,
+      "step": 22210
+    },
+    {
+      "epoch": 0.14618421052631578,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.03775634765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1253,
+      "loss/crossentropy": 2.4262067079544067,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23887630701065063,
+      "loss/reg": 0.0,
+      "step": 22220
+    },
+    {
+      "epoch": 0.14625,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.19719136555989583,
+      "learning_rate": 0.0001,
+      "loss": 3.1058,
+      "loss/crossentropy": 2.547460687160492,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25797063410282134,
+      "loss/reg": 0.0,
+      "step": 22230
+    },
+    {
+      "epoch": 0.1463157894736842,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.2289446512858073,
+      "learning_rate": 0.0001,
+      "loss": 3.1238,
+      "loss/crossentropy": 2.1126400232315063,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23666388988494874,
+      "loss/reg": 0.0,
+      "step": 22240
+    },
+    {
+      "epoch": 0.14638157894736842,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.20608495076497396,
+      "learning_rate": 0.0001,
+      "loss": 3.0914,
+      "loss/crossentropy": 2.5851184129714966,
+      "loss/hidden": 2.8359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2918281376361847,
+      "loss/reg": 0.0,
+      "step": 22250
+    },
+    {
+      "epoch": 0.14644736842105263,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.5136329650878906,
+      "learning_rate": 0.0001,
+      "loss": 3.0971,
+      "loss/crossentropy": 2.319287371635437,
+      "loss/hidden": 3.0140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.283975313603878,
+      "loss/reg": 0.0,
+      "step": 22260
+    },
+    {
+      "epoch": 0.14651315789473685,
+      "grad_norm": 2.625,
+      "grad_norm_var": 9.41323216756185,
+      "learning_rate": 0.0001,
+      "loss": 3.2274,
+      "loss/crossentropy": 2.350848126411438,
+      "loss/hidden": 2.884375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3113620936870575,
+      "loss/reg": 0.0,
+      "step": 22270
+    },
+    {
+      "epoch": 0.14657894736842106,
+      "grad_norm": 3.71875,
+      "grad_norm_var": 0.20803120930989583,
+      "learning_rate": 0.0001,
+      "loss": 3.1285,
+      "loss/crossentropy": 2.175449311733246,
+      "loss/hidden": 2.8640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.235707426071167,
+      "loss/reg": 0.0,
+      "step": 22280
+    },
+    {
+      "epoch": 0.14664473684210527,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.37470601399739584,
+      "learning_rate": 0.0001,
+      "loss": 3.1206,
+      "loss/crossentropy": 2.6359380006790163,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2809421643614769,
+      "loss/reg": 0.0,
+      "step": 22290
+    },
+    {
+      "epoch": 0.14671052631578949,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.33459879557291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1002,
+      "loss/crossentropy": 2.2076025128364565,
+      "loss/hidden": 2.7359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22375066131353377,
+      "loss/reg": 0.0,
+      "step": 22300
+    },
+    {
+      "epoch": 0.14677631578947367,
+      "grad_norm": 4.34375,
+      "grad_norm_var": 0.37032877604166664,
+      "learning_rate": 0.0001,
+      "loss": 3.117,
+      "loss/crossentropy": 2.0601858735084533,
+      "loss/hidden": 2.971875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3449657797813416,
+      "loss/reg": 0.0,
+      "step": 22310
+    },
+    {
+      "epoch": 0.14684210526315788,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.7130360921223958,
+      "learning_rate": 0.0001,
+      "loss": 3.0669,
+      "loss/crossentropy": 2.2765709161758423,
+      "loss/hidden": 2.6734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2070981428027153,
+      "loss/reg": 0.0,
+      "step": 22320
+    },
+    {
+      "epoch": 0.1469078947368421,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.5365193684895834,
+      "learning_rate": 0.0001,
+      "loss": 3.0347,
+      "loss/crossentropy": 2.1606739163398743,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2066648006439209,
+      "loss/reg": 0.0,
+      "step": 22330
+    },
+    {
+      "epoch": 0.1469736842105263,
+      "grad_norm": 11.25,
+      "grad_norm_var": 5.021190388997396,
+      "learning_rate": 0.0001,
+      "loss": 3.1108,
+      "loss/crossentropy": 2.4670102834701537,
+      "loss/hidden": 2.621875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20643151551485062,
+      "loss/reg": 0.0,
+      "step": 22340
+    },
+    {
+      "epoch": 0.14703947368421053,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 4.964111328125,
+      "learning_rate": 0.0001,
+      "loss": 3.0792,
+      "loss/crossentropy": 2.1708640813827516,
+      "loss/hidden": 2.96875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2493487134575844,
+      "loss/reg": 0.0,
+      "step": 22350
+    },
+    {
+      "epoch": 0.14710526315789474,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.2525299072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.1371,
+      "loss/crossentropy": 2.4388205766677857,
+      "loss/hidden": 2.9765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25962347984313966,
+      "loss/reg": 0.0,
+      "step": 22360
+    },
+    {
+      "epoch": 0.14717105263157895,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.1518707275390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0391,
+      "loss/crossentropy": 2.295293319225311,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2302995279431343,
+      "loss/reg": 0.0,
+      "step": 22370
+    },
+    {
+      "epoch": 0.14723684210526317,
+      "grad_norm": 3.609375,
+      "grad_norm_var": 0.20093994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1142,
+      "loss/crossentropy": 2.081736671924591,
+      "loss/hidden": 2.7296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21390146017074585,
+      "loss/reg": 0.0,
+      "step": 22380
+    },
+    {
+      "epoch": 0.14730263157894738,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.1891021728515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0243,
+      "loss/crossentropy": 2.1469112396240235,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24780639857053757,
+      "loss/reg": 0.0,
+      "step": 22390
+    },
+    {
+      "epoch": 0.14736842105263157,
+      "grad_norm": 3.5,
+      "grad_norm_var": 0.41617431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1187,
+      "loss/crossentropy": 2.219987118244171,
+      "loss/hidden": 3.0078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2700192674994469,
+      "loss/reg": 0.0,
+      "step": 22400
+    },
+    {
+      "epoch": 0.14743421052631578,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.10216471354166666,
+      "learning_rate": 0.0001,
+      "loss": 3.1659,
+      "loss/crossentropy": 2.387179672718048,
+      "loss/hidden": 2.884375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25483787804841995,
+      "loss/reg": 0.0,
+      "step": 22410
+    },
+    {
+      "epoch": 0.1475,
+      "grad_norm": 3.21875,
+      "grad_norm_var": 0.10955403645833334,
+      "learning_rate": 0.0001,
+      "loss": 3.1812,
+      "loss/crossentropy": 2.4982651591300966,
+      "loss/hidden": 3.0203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23052496314048768,
+      "loss/reg": 0.0,
+      "step": 22420
+    },
+    {
+      "epoch": 0.1475657894736842,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.0999664306640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0656,
+      "loss/crossentropy": 2.2789443135261536,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23072773963212967,
+      "loss/reg": 0.0,
+      "step": 22430
+    },
+    {
+      "epoch": 0.14763157894736842,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.074658203125,
+      "learning_rate": 0.0001,
+      "loss": 3.0366,
+      "loss/crossentropy": 2.384058713912964,
+      "loss/hidden": 2.6921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2288333684206009,
+      "loss/reg": 0.0,
+      "step": 22440
+    },
+    {
+      "epoch": 0.14769736842105263,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.05872294108072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1249,
+      "loss/crossentropy": 2.1794327974319456,
+      "loss/hidden": 2.959375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25599117279052735,
+      "loss/reg": 0.0,
+      "step": 22450
+    },
+    {
+      "epoch": 0.14776315789473685,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.20020243326822917,
+      "learning_rate": 0.0001,
+      "loss": 3.2428,
+      "loss/crossentropy": 2.3936930537223815,
+      "loss/hidden": 3.01875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33337023556232454,
+      "loss/reg": 0.0,
+      "step": 22460
+    },
+    {
+      "epoch": 0.14782894736842106,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.2228424072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0933,
+      "loss/crossentropy": 2.290210509300232,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22466516196727754,
+      "loss/reg": 0.0,
+      "step": 22470
+    },
+    {
+      "epoch": 0.14789473684210527,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.07656962076822917,
+      "learning_rate": 0.0001,
+      "loss": 3.1112,
+      "loss/crossentropy": 2.1793927550315857,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2438851624727249,
+      "loss/reg": 0.0,
+      "step": 22480
+    },
+    {
+      "epoch": 0.1479605263157895,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.15196024576822917,
+      "learning_rate": 0.0001,
+      "loss": 3.105,
+      "loss/crossentropy": 2.3326605677604677,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22931813299655915,
+      "loss/reg": 0.0,
+      "step": 22490
+    },
+    {
+      "epoch": 0.14802631578947367,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.15309244791666668,
+      "learning_rate": 0.0001,
+      "loss": 3.0875,
+      "loss/crossentropy": 2.222768235206604,
+      "loss/hidden": 2.9828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2537077710032463,
+      "loss/reg": 0.0,
+      "step": 22500
+    },
+    {
+      "epoch": 0.14809210526315789,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.0938385009765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1419,
+      "loss/crossentropy": 2.308623898029327,
+      "loss/hidden": 3.0859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27548344135284425,
+      "loss/reg": 0.0,
+      "step": 22510
+    },
+    {
+      "epoch": 0.1481578947368421,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.07166239420572916,
+      "learning_rate": 0.0001,
+      "loss": 3.068,
+      "loss/crossentropy": 2.261019694805145,
+      "loss/hidden": 2.765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23012840151786804,
+      "loss/reg": 0.0,
+      "step": 22520
+    },
+    {
+      "epoch": 0.1482236842105263,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.13349202473958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1074,
+      "loss/crossentropy": 2.265119397640228,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2193696081638336,
+      "loss/reg": 0.0,
+      "step": 22530
+    },
+    {
+      "epoch": 0.14828947368421053,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 2.090127618074695e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.3232,
+      "loss/crossentropy": 2.414242672920227,
+      "loss/hidden": 2.78125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24838838130235671,
+      "loss/reg": 0.0,
+      "step": 22540
+    },
+    {
+      "epoch": 0.14835526315789474,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.0516265869140625,
+      "learning_rate": 0.0001,
+      "loss": 3.147,
+      "loss/crossentropy": 2.4570279359817504,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27158954441547395,
+      "loss/reg": 0.0,
+      "step": 22550
+    },
+    {
+      "epoch": 0.14842105263157895,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.0767229715983073,
+      "learning_rate": 0.0001,
+      "loss": 3.0495,
+      "loss/crossentropy": 2.1562333941459655,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.252650773525238,
+      "loss/reg": 0.0,
+      "step": 22560
+    },
+    {
+      "epoch": 0.14848684210526317,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.2514625549316406,
+      "learning_rate": 0.0001,
+      "loss": 3.1717,
+      "loss/crossentropy": 2.0726990699768066,
+      "loss/hidden": 3.05,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2678316295146942,
+      "loss/reg": 0.0,
+      "step": 22570
+    },
+    {
+      "epoch": 0.14855263157894738,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.21327718098958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1387,
+      "loss/crossentropy": 2.32381352186203,
+      "loss/hidden": 2.9,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2564924195408821,
+      "loss/reg": 0.0,
+      "step": 22580
+    },
+    {
+      "epoch": 0.14861842105263157,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.08567301432291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1625,
+      "loss/crossentropy": 2.15835280418396,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3166464313864708,
+      "loss/reg": 0.0,
+      "step": 22590
+    },
+    {
+      "epoch": 0.14868421052631578,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.13315327962239584,
+      "learning_rate": 0.0001,
+      "loss": 3.0573,
+      "loss/crossentropy": 2.5024718761444094,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2705509811639786,
+      "loss/reg": 0.0,
+      "step": 22600
+    },
+    {
+      "epoch": 0.14875,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.1475176493326823,
+      "learning_rate": 0.0001,
+      "loss": 3.0119,
+      "loss/crossentropy": 2.2811665177345275,
+      "loss/hidden": 2.85625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23891925811767578,
+      "loss/reg": 0.0,
+      "step": 22610
+    },
+    {
+      "epoch": 0.1488157894736842,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 31.435579172770183,
+      "learning_rate": 0.0001,
+      "loss": 3.0136,
+      "loss/crossentropy": 2.1407122373580934,
+      "loss/hidden": 2.9484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26466352492570877,
+      "loss/reg": 0.0,
+      "step": 22620
+    },
+    {
+      "epoch": 0.14888157894736842,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.10041402180989584,
+      "learning_rate": 0.0001,
+      "loss": 3.0561,
+      "loss/crossentropy": 2.3511832118034364,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26619747579097747,
+      "loss/reg": 0.0,
+      "step": 22630
+    },
+    {
+      "epoch": 0.14894736842105263,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.16308186848958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1069,
+      "loss/crossentropy": 2.1790146827697754,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2223970666527748,
+      "loss/reg": 0.0,
+      "step": 22640
+    },
+    {
+      "epoch": 0.14901315789473685,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.48355204264322915,
+      "learning_rate": 0.0001,
+      "loss": 3.1746,
+      "loss/crossentropy": 2.2969871520996095,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28595788925886156,
+      "loss/reg": 0.0,
+      "step": 22650
+    },
+    {
+      "epoch": 0.14907894736842106,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.38804931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1378,
+      "loss/crossentropy": 2.484911561012268,
+      "loss/hidden": 2.975,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28142704218626025,
+      "loss/reg": 0.0,
+      "step": 22660
+    },
+    {
+      "epoch": 0.14914473684210527,
+      "grad_norm": 3.359375,
+      "grad_norm_var": 0.11812744140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2449,
+      "loss/crossentropy": 2.3331188917160035,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2567854106426239,
+      "loss/reg": 0.0,
+      "step": 22670
+    },
+    {
+      "epoch": 0.14921052631578946,
+      "grad_norm": 2.0,
+      "grad_norm_var": 0.11311848958333333,
+      "learning_rate": 0.0001,
+      "loss": 3.0464,
+      "loss/crossentropy": 2.5099231958389283,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22425275444984435,
+      "loss/reg": 0.0,
+      "step": 22680
+    },
+    {
+      "epoch": 0.14927631578947367,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.033707682291666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0841,
+      "loss/crossentropy": 2.5707973539829254,
+      "loss/hidden": 2.971875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27995080798864364,
+      "loss/reg": 0.0,
+      "step": 22690
+    },
+    {
+      "epoch": 0.1493421052631579,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.03998921712239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0877,
+      "loss/crossentropy": 2.2857648015022276,
+      "loss/hidden": 2.8140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2597381517291069,
+      "loss/reg": 0.0,
+      "step": 22700
+    },
+    {
+      "epoch": 0.1494078947368421,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.05201416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0551,
+      "loss/crossentropy": 2.4230233311653135,
+      "loss/hidden": 2.93125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2558469220995903,
+      "loss/reg": 0.0,
+      "step": 22710
+    },
+    {
+      "epoch": 0.14947368421052631,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.08981831868489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0629,
+      "loss/crossentropy": 2.3983253479003905,
+      "loss/hidden": 2.753125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25465885996818544,
+      "loss/reg": 0.0,
+      "step": 22720
+    },
+    {
+      "epoch": 0.14953947368421053,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.14431966145833333,
+      "learning_rate": 0.0001,
+      "loss": 3.1884,
+      "loss/crossentropy": 2.227739405632019,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27533823624253273,
+      "loss/reg": 0.0,
+      "step": 22730
+    },
+    {
+      "epoch": 0.14960526315789474,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.09478759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.044,
+      "loss/crossentropy": 2.2618382215499877,
+      "loss/hidden": 2.7265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22343547642230988,
+      "loss/reg": 0.0,
+      "step": 22740
+    },
+    {
+      "epoch": 0.14967105263157895,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.18606363932291667,
+      "learning_rate": 0.0001,
+      "loss": 3.145,
+      "loss/crossentropy": 2.4264959335327148,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26770799309015275,
+      "loss/reg": 0.0,
+      "step": 22750
+    },
+    {
+      "epoch": 0.14973684210526317,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.0256500244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.2178,
+      "loss/crossentropy": 2.603011679649353,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28421192467212675,
+      "loss/reg": 0.0,
+      "step": 22760
+    },
+    {
+      "epoch": 0.14980263157894738,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.04109598795572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0566,
+      "loss/crossentropy": 2.239471447467804,
+      "loss/hidden": 2.815625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2365947112441063,
+      "loss/reg": 0.0,
+      "step": 22770
+    },
+    {
+      "epoch": 0.14986842105263157,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.11721903483072917,
+      "learning_rate": 0.0001,
+      "loss": 3.0489,
+      "loss/crossentropy": 2.264164900779724,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2334253668785095,
+      "loss/reg": 0.0,
+      "step": 22780
+    },
+    {
+      "epoch": 0.14993421052631578,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.1690093994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1266,
+      "loss/crossentropy": 2.3352009534835814,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23781208097934722,
+      "loss/reg": 0.0,
+      "step": 22790
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 3.25,
+      "grad_norm_var": 0.25569559733072916,
+      "learning_rate": 0.0001,
+      "loss": 3.1863,
+      "loss/crossentropy": 2.260995364189148,
+      "loss/hidden": 2.9921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28239033967256544,
+      "loss/reg": 0.0,
+      "step": 22800
+    },
+    {
+      "epoch": 0.1500657894736842,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.5896321614583333,
+      "learning_rate": 0.0001,
+      "loss": 3.1214,
+      "loss/crossentropy": 2.487893545627594,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2422095239162445,
+      "loss/reg": 0.0,
+      "step": 22810
+    },
+    {
+      "epoch": 0.15013157894736842,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.44045817057291664,
+      "learning_rate": 0.0001,
+      "loss": 3.0838,
+      "loss/crossentropy": 2.484136939048767,
+      "loss/hidden": 2.7140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22923200577497482,
+      "loss/reg": 0.0,
+      "step": 22820
+    },
+    {
+      "epoch": 0.15019736842105263,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 3.111881782327837e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2029,
+      "loss/crossentropy": 2.462354898452759,
+      "loss/hidden": 2.7390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23231214433908462,
+      "loss/reg": 0.0,
+      "step": 22830
+    },
+    {
+      "epoch": 0.15026315789473685,
+      "grad_norm": 2.25,
+      "grad_norm_var": 3.1118817821970925e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1305,
+      "loss/crossentropy": 2.1894614577293394,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24518598318099977,
+      "loss/reg": 0.0,
+      "step": 22840
+    },
+    {
+      "epoch": 0.15032894736842106,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.23975321451822917,
+      "learning_rate": 0.0001,
+      "loss": 3.0371,
+      "loss/crossentropy": 2.427228879928589,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.250908388197422,
+      "loss/reg": 0.0,
+      "step": 22850
+    },
+    {
+      "epoch": 0.15039473684210528,
+      "grad_norm": 7.125,
+      "grad_norm_var": 1.4729563395182292,
+      "learning_rate": 0.0001,
+      "loss": 3.1606,
+      "loss/crossentropy": 2.163099730014801,
+      "loss/hidden": 3.0328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2531519740819931,
+      "loss/reg": 0.0,
+      "step": 22860
+    },
+    {
+      "epoch": 0.15046052631578946,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 1.7556060791015624,
+      "learning_rate": 0.0001,
+      "loss": 3.127,
+      "loss/crossentropy": 2.4134068608284,
+      "loss/hidden": 2.7875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2419949784874916,
+      "loss/reg": 0.0,
+      "step": 22870
+    },
+    {
+      "epoch": 0.15052631578947367,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.6191080729166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0458,
+      "loss/crossentropy": 2.2460831999778748,
+      "loss/hidden": 2.8625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25571899116039276,
+      "loss/reg": 0.0,
+      "step": 22880
+    },
+    {
+      "epoch": 0.1505921052631579,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.1325592041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1666,
+      "loss/crossentropy": 2.32741322517395,
+      "loss/hidden": 2.840625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23611859530210494,
+      "loss/reg": 0.0,
+      "step": 22890
+    },
+    {
+      "epoch": 0.1506578947368421,
+      "grad_norm": 4.03125,
+      "grad_norm_var": 0.24289957682291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1655,
+      "loss/crossentropy": 2.436409044265747,
+      "loss/hidden": 3.04375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28332450836896894,
+      "loss/reg": 0.0,
+      "step": 22900
+    },
+    {
+      "epoch": 0.15072368421052632,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.23401590983072917,
+      "learning_rate": 0.0001,
+      "loss": 2.9932,
+      "loss/crossentropy": 2.2307902753353117,
+      "loss/hidden": 2.9,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23847298920154572,
+      "loss/reg": 0.0,
+      "step": 22910
+    },
+    {
+      "epoch": 0.15078947368421053,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.12184244791666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1045,
+      "loss/crossentropy": 2.2170926928520203,
+      "loss/hidden": 2.8984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23038018941879274,
+      "loss/reg": 0.0,
+      "step": 22920
+    },
+    {
+      "epoch": 0.15085526315789474,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.09688695271809895,
+      "learning_rate": 0.0001,
+      "loss": 3.0845,
+      "loss/crossentropy": 2.1675124526023866,
+      "loss/hidden": 2.8203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20779158547520638,
+      "loss/reg": 0.0,
+      "step": 22930
+    },
+    {
+      "epoch": 0.15092105263157896,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.7031776428222656,
+      "learning_rate": 0.0001,
+      "loss": 3.131,
+      "loss/crossentropy": 2.483162760734558,
+      "loss/hidden": 2.9015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2888242840766907,
+      "loss/reg": 0.0,
+      "step": 22940
+    },
+    {
+      "epoch": 0.15098684210526317,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.6901682535807292,
+      "learning_rate": 0.0001,
+      "loss": 3.1206,
+      "loss/crossentropy": 1.9849469184875488,
+      "loss/hidden": 2.8640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21038227528333664,
+      "loss/reg": 0.0,
+      "step": 22950
+    },
+    {
+      "epoch": 0.15105263157894736,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.08320210774739584,
+      "learning_rate": 0.0001,
+      "loss": 3.1315,
+      "loss/crossentropy": 2.424567532539368,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2791455164551735,
+      "loss/reg": 0.0,
+      "step": 22960
+    },
+    {
+      "epoch": 0.15111842105263157,
+      "grad_norm": 2.0,
+      "grad_norm_var": 0.0634674072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.054,
+      "loss/crossentropy": 2.531536269187927,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24857166707515715,
+      "loss/reg": 0.0,
+      "step": 22970
+    },
+    {
+      "epoch": 0.15118421052631578,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.06180013020833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0542,
+      "loss/crossentropy": 2.4494295597076414,
+      "loss/hidden": 2.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24070152640342712,
+      "loss/reg": 0.0,
+      "step": 22980
+    },
+    {
+      "epoch": 0.15125,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.07551167805989584,
+      "learning_rate": 0.0001,
+      "loss": 3.0773,
+      "loss/crossentropy": 2.418122172355652,
+      "loss/hidden": 2.740625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24059886187314988,
+      "loss/reg": 0.0,
+      "step": 22990
+    },
+    {
+      "epoch": 0.1513157894736842,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.1075592041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1222,
+      "loss/crossentropy": 2.184150278568268,
+      "loss/hidden": 2.94375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26524696350097654,
+      "loss/reg": 0.0,
+      "step": 23000
+    },
+    {
+      "epoch": 0.15138157894736842,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.12967020670572918,
+      "learning_rate": 0.0001,
+      "loss": 3.043,
+      "loss/crossentropy": 2.48363002538681,
+      "loss/hidden": 2.709375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22335316240787506,
+      "loss/reg": 0.0,
+      "step": 23010
+    },
+    {
+      "epoch": 0.15144736842105264,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.1149810791015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1275,
+      "loss/crossentropy": 2.305170452594757,
+      "loss/hidden": 2.7234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23285460025072097,
+      "loss/reg": 0.0,
+      "step": 23020
+    },
+    {
+      "epoch": 0.15151315789473685,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.06519775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0832,
+      "loss/crossentropy": 2.5363330364227297,
+      "loss/hidden": 2.7359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22520080506801604,
+      "loss/reg": 0.0,
+      "step": 23030
+    },
+    {
+      "epoch": 0.15157894736842106,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.39241129557291665,
+      "learning_rate": 0.0001,
+      "loss": 3.113,
+      "loss/crossentropy": 2.1367889642715454,
+      "loss/hidden": 2.7390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29226300269365313,
+      "loss/reg": 0.0,
+      "step": 23040
+    },
+    {
+      "epoch": 0.15164473684210528,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.39695638020833335,
+      "learning_rate": 0.0001,
+      "loss": 3.1147,
+      "loss/crossentropy": 2.4076698064804076,
+      "loss/hidden": 2.84375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34615216627717016,
+      "loss/reg": 0.0,
+      "step": 23050
+    },
+    {
+      "epoch": 0.15171052631578946,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.23273824055989584,
+      "learning_rate": 0.0001,
+      "loss": 3.0922,
+      "loss/crossentropy": 2.2620026230812074,
+      "loss/hidden": 2.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24059030711650847,
+      "loss/reg": 0.0,
+      "step": 23060
+    },
+    {
+      "epoch": 0.15177631578947368,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.258984375,
+      "learning_rate": 0.0001,
+      "loss": 3.0849,
+      "loss/crossentropy": 2.4243380606174467,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28078764528036115,
+      "loss/reg": 0.0,
+      "step": 23070
+    },
+    {
+      "epoch": 0.1518421052631579,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 4.336297930295782e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1967,
+      "loss/crossentropy": 2.3340250134468077,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2355951637029648,
+      "loss/reg": 0.0,
+      "step": 23080
+    },
+    {
+      "epoch": 0.1519078947368421,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 4.336297930357517e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.031,
+      "loss/crossentropy": 2.1574897170066833,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21495652794837952,
+      "loss/reg": 0.0,
+      "step": 23090
+    },
+    {
+      "epoch": 0.15197368421052632,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.1123199462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1392,
+      "loss/crossentropy": 2.405716967582703,
+      "loss/hidden": 2.8671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27174981236457824,
+      "loss/reg": 0.0,
+      "step": 23100
+    },
+    {
+      "epoch": 0.15203947368421053,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 2.003859923676365e+16,
+      "learning_rate": 0.0001,
+      "loss": 3.2336,
+      "loss/crossentropy": 2.339055967330933,
+      "loss/hidden": 3.0265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28572248220443724,
+      "loss/reg": 0.0,
+      "step": 23110
+    },
+    {
+      "epoch": 0.15210526315789474,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.21015218098958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0005,
+      "loss/crossentropy": 2.2984538078308105,
+      "loss/hidden": 2.684375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21156432926654817,
+      "loss/reg": 0.0,
+      "step": 23120
+    },
+    {
+      "epoch": 0.15217105263157896,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.08379618326822917,
+      "learning_rate": 0.0001,
+      "loss": 3.116,
+      "loss/crossentropy": 2.3342798829078673,
+      "loss/hidden": 2.7984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22476595863699914,
+      "loss/reg": 0.0,
+      "step": 23130
+    },
+    {
+      "epoch": 0.15223684210526317,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.0606842041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0265,
+      "loss/crossentropy": 2.3210121750831605,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23322480022907258,
+      "loss/reg": 0.0,
+      "step": 23140
+    },
+    {
+      "epoch": 0.15230263157894736,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.23906962076822916,
+      "learning_rate": 0.0001,
+      "loss": 3.0813,
+      "loss/crossentropy": 2.391731929779053,
+      "loss/hidden": 3.05625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3979014977812767,
+      "loss/reg": 0.0,
+      "step": 23150
+    },
+    {
+      "epoch": 0.15236842105263157,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.049214680989583336,
+      "learning_rate": 0.0001,
+      "loss": 3.0574,
+      "loss/crossentropy": 2.492924761772156,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23957584500312806,
+      "loss/reg": 0.0,
+      "step": 23160
+    },
+    {
+      "epoch": 0.15243421052631578,
+      "grad_norm": 1.9765625,
+      "grad_norm_var": 0.07945938110351562,
+      "learning_rate": 0.0001,
+      "loss": 3.0317,
+      "loss/crossentropy": 2.3342637419700623,
+      "loss/hidden": 2.8625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30051169246435167,
+      "loss/reg": 0.0,
+      "step": 23170
+    },
+    {
+      "epoch": 0.1525,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.1088396708170573,
+      "learning_rate": 0.0001,
+      "loss": 3.1944,
+      "loss/crossentropy": 2.306346225738525,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.246099853515625,
+      "loss/reg": 0.0,
+      "step": 23180
+    },
+    {
+      "epoch": 0.1525657894736842,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.17786839803059895,
+      "learning_rate": 0.0001,
+      "loss": 3.0341,
+      "loss/crossentropy": 2.030124247074127,
+      "loss/hidden": 2.9328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21610897928476333,
+      "loss/reg": 0.0,
+      "step": 23190
+    },
+    {
+      "epoch": 0.15263157894736842,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.12164688110351562,
+      "learning_rate": 0.0001,
+      "loss": 3.0381,
+      "loss/crossentropy": 2.3866997003555297,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25531851798295974,
+      "loss/reg": 0.0,
+      "step": 23200
+    },
+    {
+      "epoch": 0.15269736842105264,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.13189697265625,
+      "learning_rate": 0.0001,
+      "loss": 3.1059,
+      "loss/crossentropy": 2.6614113092422484,
+      "loss/hidden": 2.840625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26844717264175416,
+      "loss/reg": 0.0,
+      "step": 23210
+    },
+    {
+      "epoch": 0.15276315789473685,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.24109598795572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0657,
+      "loss/crossentropy": 2.384077048301697,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2543630987405777,
+      "loss/reg": 0.0,
+      "step": 23220
+    },
+    {
+      "epoch": 0.15282894736842106,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.27134501139322914,
+      "learning_rate": 0.0001,
+      "loss": 3.0512,
+      "loss/crossentropy": 2.4361581921577455,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24168266355991364,
+      "loss/reg": 0.0,
+      "step": 23230
+    },
+    {
+      "epoch": 0.15289473684210525,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.06411031087239584,
+      "learning_rate": 0.0001,
+      "loss": 3.1163,
+      "loss/crossentropy": 2.439437985420227,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23927776366472245,
+      "loss/reg": 0.0,
+      "step": 23240
+    },
+    {
+      "epoch": 0.15296052631578946,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.20168355305989583,
+      "learning_rate": 0.0001,
+      "loss": 3.1405,
+      "loss/crossentropy": 1.9147474735975265,
+      "loss/hidden": 2.9296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27681511342525483,
+      "loss/reg": 0.0,
+      "step": 23250
+    },
+    {
+      "epoch": 0.15302631578947368,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.3036122639973958,
+      "learning_rate": 0.0001,
+      "loss": 3.1139,
+      "loss/crossentropy": 1.9262932538986206,
+      "loss/hidden": 2.9421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2078495942056179,
+      "loss/reg": 0.0,
+      "step": 23260
+    },
+    {
+      "epoch": 0.1530921052631579,
+      "grad_norm": 3.1875,
+      "grad_norm_var": 0.26516520182291664,
+      "learning_rate": 0.0001,
+      "loss": 3.0706,
+      "loss/crossentropy": 2.252467918395996,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24831898286938667,
+      "loss/reg": 0.0,
+      "step": 23270
+    },
+    {
+      "epoch": 0.1531578947368421,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.12827860514322917,
+      "learning_rate": 0.0001,
+      "loss": 3.1631,
+      "loss/crossentropy": 2.0835241615772246,
+      "loss/hidden": 3.0796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2629551820456982,
+      "loss/reg": 0.0,
+      "step": 23280
+    },
+    {
+      "epoch": 0.15322368421052632,
+      "grad_norm": 4.0625,
+      "grad_norm_var": 0.2850494384765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0116,
+      "loss/crossentropy": 2.533276152610779,
+      "loss/hidden": 2.6890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25189711451530455,
+      "loss/reg": 0.0,
+      "step": 23290
+    },
+    {
+      "epoch": 0.15328947368421053,
+      "grad_norm": 1.984375,
+      "grad_norm_var": 0.28486328125,
+      "learning_rate": 0.0001,
+      "loss": 3.0549,
+      "loss/crossentropy": 2.2883153557777405,
+      "loss/hidden": 2.6359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20354210436344147,
+      "loss/reg": 0.0,
+      "step": 23300
+    },
+    {
+      "epoch": 0.15335526315789474,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.1171295166015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1273,
+      "loss/crossentropy": 2.3346426010131838,
+      "loss/hidden": 3.0015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32933240532875063,
+      "loss/reg": 0.0,
+      "step": 23310
+    },
+    {
+      "epoch": 0.15342105263157896,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.15787760416666666,
+      "learning_rate": 0.0001,
+      "loss": 3.141,
+      "loss/crossentropy": 2.1821151852607725,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2542765408754349,
+      "loss/reg": 0.0,
+      "step": 23320
+    },
+    {
+      "epoch": 0.15348684210526317,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.1633453369140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1175,
+      "loss/crossentropy": 2.303962028026581,
+      "loss/hidden": 3.0,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25389211922883986,
+      "loss/reg": 0.0,
+      "step": 23330
+    },
+    {
+      "epoch": 0.15355263157894736,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 6.715729777018229,
+      "learning_rate": 0.0001,
+      "loss": 3.1862,
+      "loss/crossentropy": 2.387230467796326,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26721241027116777,
+      "loss/reg": 0.0,
+      "step": 23340
+    },
+    {
+      "epoch": 0.15361842105263157,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.14257405598958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1622,
+      "loss/crossentropy": 2.4242820143699646,
+      "loss/hidden": 2.8671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29456254839897156,
+      "loss/reg": 0.0,
+      "step": 23350
+    },
+    {
+      "epoch": 0.15368421052631578,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.18007405598958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1082,
+      "loss/crossentropy": 2.1694801568984987,
+      "loss/hidden": 2.9765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3005404189229012,
+      "loss/reg": 0.0,
+      "step": 23360
+    },
+    {
+      "epoch": 0.15375,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.13083089192708333,
+      "learning_rate": 0.0001,
+      "loss": 3.0146,
+      "loss/crossentropy": 2.5201270818710326,
+      "loss/hidden": 2.8,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23150605708360672,
+      "loss/reg": 0.0,
+      "step": 23370
+    },
+    {
+      "epoch": 0.1538157894736842,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.1706621805826823,
+      "learning_rate": 0.0001,
+      "loss": 3.04,
+      "loss/crossentropy": 2.4117377281188963,
+      "loss/hidden": 2.6984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22549189925193786,
+      "loss/reg": 0.0,
+      "step": 23380
+    },
+    {
+      "epoch": 0.15388157894736842,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.049843088785807295,
+      "learning_rate": 0.0001,
+      "loss": 3.0355,
+      "loss/crossentropy": 2.166217362880707,
+      "loss/hidden": 2.8453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2460268869996071,
+      "loss/reg": 0.0,
+      "step": 23390
+    },
+    {
+      "epoch": 0.15394736842105264,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.027437337239583335,
+      "learning_rate": 0.0001,
+      "loss": 3.0332,
+      "loss/crossentropy": 2.2912797331809998,
+      "loss/hidden": 2.671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2184738412499428,
+      "loss/reg": 0.0,
+      "step": 23400
+    },
+    {
+      "epoch": 0.15401315789473685,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.02008056640625,
+      "learning_rate": 0.0001,
+      "loss": 3.049,
+      "loss/crossentropy": 2.2143723726272584,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23227078467607498,
+      "loss/reg": 0.0,
+      "step": 23410
+    },
+    {
+      "epoch": 0.15407894736842107,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.04762140909830729,
+      "learning_rate": 0.0001,
+      "loss": 2.9746,
+      "loss/crossentropy": 2.1996548891067507,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20859691947698594,
+      "loss/reg": 0.0,
+      "step": 23420
+    },
+    {
+      "epoch": 0.15414473684210525,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.08440348307291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1332,
+      "loss/crossentropy": 2.353720319271088,
+      "loss/hidden": 3.0734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32487219721078875,
+      "loss/reg": 0.0,
+      "step": 23430
+    },
+    {
+      "epoch": 0.15421052631578946,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.04627278645833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0352,
+      "loss/crossentropy": 1.994243037700653,
+      "loss/hidden": 2.7453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2228219524025917,
+      "loss/reg": 0.0,
+      "step": 23440
+    },
+    {
+      "epoch": 0.15427631578947368,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.025699869791666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0571,
+      "loss/crossentropy": 2.368081831932068,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24828383028507234,
+      "loss/reg": 0.0,
+      "step": 23450
+    },
+    {
+      "epoch": 0.1543421052631579,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.06443684895833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0708,
+      "loss/crossentropy": 2.1184537053108214,
+      "loss/hidden": 3.021875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2568272680044174,
+      "loss/reg": 0.0,
+      "step": 23460
+    },
+    {
+      "epoch": 0.1544078947368421,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.13352762858072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1323,
+      "loss/crossentropy": 2.3327906489372254,
+      "loss/hidden": 2.9,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2940856009721756,
+      "loss/reg": 0.0,
+      "step": 23470
+    },
+    {
+      "epoch": 0.15447368421052632,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.14630533854166666,
+      "learning_rate": 0.0001,
+      "loss": 3.0207,
+      "loss/crossentropy": 2.5257789850234986,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24059889316558838,
+      "loss/reg": 0.0,
+      "step": 23480
+    },
+    {
+      "epoch": 0.15453947368421053,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.059357706705729166,
+      "learning_rate": 0.0001,
+      "loss": 3.0888,
+      "loss/crossentropy": 2.0001726031303404,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2182185634970665,
+      "loss/reg": 0.0,
+      "step": 23490
+    },
+    {
+      "epoch": 0.15460526315789475,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.07535171508789062,
+      "learning_rate": 0.0001,
+      "loss": 3.0415,
+      "loss/crossentropy": 2.4137069463729857,
+      "loss/hidden": 2.809375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26817760467529295,
+      "loss/reg": 0.0,
+      "step": 23500
+    },
+    {
+      "epoch": 0.15467105263157896,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.09633967081705729,
+      "learning_rate": 0.0001,
+      "loss": 3.0427,
+      "loss/crossentropy": 2.534043550491333,
+      "loss/hidden": 2.63125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22827706933021547,
+      "loss/reg": 0.0,
+      "step": 23510
+    },
+    {
+      "epoch": 0.15473684210526314,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.0346832275390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0819,
+      "loss/crossentropy": 2.1481072187423704,
+      "loss/hidden": 2.91875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22196744605898858,
+      "loss/reg": 0.0,
+      "step": 23520
+    },
+    {
+      "epoch": 0.15480263157894736,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.07886962890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1167,
+      "loss/crossentropy": 2.0396682798862455,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22266072854399682,
+      "loss/reg": 0.0,
+      "step": 23530
+    },
+    {
+      "epoch": 0.15486842105263157,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.038182576497395836,
+      "learning_rate": 0.0001,
+      "loss": 3.075,
+      "loss/crossentropy": 2.164088273048401,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2313580572605133,
+      "loss/reg": 0.0,
+      "step": 23540
+    },
+    {
+      "epoch": 0.15493421052631579,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.2714515686035156,
+      "learning_rate": 0.0001,
+      "loss": 3.0593,
+      "loss/crossentropy": 2.304581105709076,
+      "loss/hidden": 2.75625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24230952113866805,
+      "loss/reg": 0.0,
+      "step": 23550
+    },
+    {
+      "epoch": 0.155,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.1161376953125,
+      "learning_rate": 0.0001,
+      "loss": 2.9886,
+      "loss/crossentropy": 1.9074640274047852,
+      "loss/hidden": 2.934375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1951706364750862,
+      "loss/reg": 0.0,
+      "step": 23560
+    },
+    {
+      "epoch": 0.1550657894736842,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.12790908813476562,
+      "learning_rate": 0.0001,
+      "loss": 3.0776,
+      "loss/crossentropy": 2.3702269911766054,
+      "loss/hidden": 3.0203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2716818228363991,
+      "loss/reg": 0.0,
+      "step": 23570
+    },
+    {
+      "epoch": 0.15513157894736843,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.1339019775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0873,
+      "loss/crossentropy": 2.205389070510864,
+      "loss/hidden": 2.921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2579692542552948,
+      "loss/reg": 0.0,
+      "step": 23580
+    },
+    {
+      "epoch": 0.15519736842105264,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.06905924479166667,
+      "learning_rate": 0.0001,
+      "loss": 3.088,
+      "loss/crossentropy": 2.091149592399597,
+      "loss/hidden": 2.8265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2086069330573082,
+      "loss/reg": 0.0,
+      "step": 23590
+    },
+    {
+      "epoch": 0.15526315789473685,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.056722005208333336,
+      "learning_rate": 0.0001,
+      "loss": 3.1228,
+      "loss/crossentropy": 2.2192450404167174,
+      "loss/hidden": 3.028125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28743477165699005,
+      "loss/reg": 0.0,
+      "step": 23600
+    },
+    {
+      "epoch": 0.15532894736842107,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.05885009765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0878,
+      "loss/crossentropy": 2.531976878643036,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2344336360692978,
+      "loss/reg": 0.0,
+      "step": 23610
+    },
+    {
+      "epoch": 0.15539473684210525,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 0.07700907389322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0331,
+      "loss/crossentropy": 2.2710886120796205,
+      "loss/hidden": 2.7390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2305494710803032,
+      "loss/reg": 0.0,
+      "step": 23620
+    },
+    {
+      "epoch": 0.15546052631578947,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.055964152018229164,
+      "learning_rate": 0.0001,
+      "loss": 3.1336,
+      "loss/crossentropy": 2.3865880966186523,
+      "loss/hidden": 2.8953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25577795803546904,
+      "loss/reg": 0.0,
+      "step": 23630
+    },
+    {
+      "epoch": 0.15552631578947368,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.05320536295572917,
+      "learning_rate": 0.0001,
+      "loss": 3.1079,
+      "loss/crossentropy": 2.388904869556427,
+      "loss/hidden": 2.921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26212394386529925,
+      "loss/reg": 0.0,
+      "step": 23640
+    },
+    {
+      "epoch": 0.1555921052631579,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.28601252237955727,
+      "learning_rate": 0.0001,
+      "loss": 3.0189,
+      "loss/crossentropy": 2.3917470812797545,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28409580439329146,
+      "loss/reg": 0.0,
+      "step": 23650
+    },
+    {
+      "epoch": 0.1556578947368421,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.05244038899739583,
+      "learning_rate": 0.0001,
+      "loss": 3.0892,
+      "loss/crossentropy": 2.266455078125,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23892848715186119,
+      "loss/reg": 0.0,
+      "step": 23660
+    },
+    {
+      "epoch": 0.15572368421052632,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.021256510416666666,
+      "learning_rate": 0.0001,
+      "loss": 3.1403,
+      "loss/crossentropy": 2.3996535420417784,
+      "loss/hidden": 3.0,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32787969559431074,
+      "loss/reg": 0.0,
+      "step": 23670
+    },
+    {
+      "epoch": 0.15578947368421053,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.03367411295572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0311,
+      "loss/crossentropy": 2.2268054604530336,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25249754935503005,
+      "loss/reg": 0.0,
+      "step": 23680
+    },
+    {
+      "epoch": 0.15585526315789475,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.112744140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1572,
+      "loss/crossentropy": 2.4530532598495483,
+      "loss/hidden": 2.971875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2573227033019066,
+      "loss/reg": 0.0,
+      "step": 23690
+    },
+    {
+      "epoch": 0.15592105263157896,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.2130767822265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0932,
+      "loss/crossentropy": 2.4343389749526976,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26323390007019043,
+      "loss/reg": 0.0,
+      "step": 23700
+    },
+    {
+      "epoch": 0.15598684210526315,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.15120340983072916,
+      "learning_rate": 0.0001,
+      "loss": 3.1834,
+      "loss/crossentropy": 2.1292245745658875,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23169804438948632,
+      "loss/reg": 0.0,
+      "step": 23710
+    },
+    {
+      "epoch": 0.15605263157894736,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.0550201416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0296,
+      "loss/crossentropy": 2.4026230216026305,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24430365711450577,
+      "loss/reg": 0.0,
+      "step": 23720
+    },
+    {
+      "epoch": 0.15611842105263157,
+      "grad_norm": 1.96875,
+      "grad_norm_var": 0.07566731770833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0245,
+      "loss/crossentropy": 2.186066722869873,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2246588721871376,
+      "loss/reg": 0.0,
+      "step": 23730
+    },
+    {
+      "epoch": 0.1561842105263158,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.05691731770833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0638,
+      "loss/crossentropy": 2.3755934476852416,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23479508757591247,
+      "loss/reg": 0.0,
+      "step": 23740
+    },
+    {
+      "epoch": 0.15625,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.06516520182291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1125,
+      "loss/crossentropy": 2.1289533019065856,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22816545218229295,
+      "loss/reg": 0.0,
+      "step": 23750
+    },
+    {
+      "epoch": 0.1563157894736842,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.15618489583333334,
+      "learning_rate": 0.0001,
+      "loss": 3.1231,
+      "loss/crossentropy": 2.5586398363113405,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2851672574877739,
+      "loss/reg": 0.0,
+      "step": 23760
+    },
+    {
+      "epoch": 0.15638157894736843,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.1464752197265625,
+      "learning_rate": 0.0001,
+      "loss": 3.063,
+      "loss/crossentropy": 2.3234678387641905,
+      "loss/hidden": 2.8796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2802169814705849,
+      "loss/reg": 0.0,
+      "step": 23770
+    },
+    {
+      "epoch": 0.15644736842105264,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 3.3502559026353274e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.217,
+      "loss/crossentropy": 2.1664799213409425,
+      "loss/hidden": 2.8765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24028320759534835,
+      "loss/reg": 0.0,
+      "step": 23780
+    },
+    {
+      "epoch": 0.15651315789473685,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 3.35025590324881e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0729,
+      "loss/crossentropy": 2.4900641202926637,
+      "loss/hidden": 2.6640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22242969423532485,
+      "loss/reg": 0.0,
+      "step": 23790
+    },
+    {
+      "epoch": 0.15657894736842104,
+      "grad_norm": 3.125,
+      "grad_norm_var": 0.42876561482747394,
+      "learning_rate": 0.0001,
+      "loss": 3.0548,
+      "loss/crossentropy": 2.285599112510681,
+      "loss/hidden": 2.9234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27018204629421233,
+      "loss/reg": 0.0,
+      "step": 23800
+    },
+    {
+      "epoch": 0.15664473684210525,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.41369527180989585,
+      "learning_rate": 0.0001,
+      "loss": 2.9886,
+      "loss/crossentropy": 2.352921783924103,
+      "loss/hidden": 2.659375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20102634727954866,
+      "loss/reg": 0.0,
+      "step": 23810
+    },
+    {
+      "epoch": 0.15671052631578947,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.11093343098958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0036,
+      "loss/crossentropy": 2.4093087553977965,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24685168713331224,
+      "loss/reg": 0.0,
+      "step": 23820
+    },
+    {
+      "epoch": 0.15677631578947368,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.0461578369140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0504,
+      "loss/crossentropy": 2.2815535068511963,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21896685659885406,
+      "loss/reg": 0.0,
+      "step": 23830
+    },
+    {
+      "epoch": 0.1568421052631579,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.07251561482747396,
+      "learning_rate": 0.0001,
+      "loss": 2.9885,
+      "loss/crossentropy": 2.6134737491607667,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22691430598497392,
+      "loss/reg": 0.0,
+      "step": 23840
+    },
+    {
+      "epoch": 0.1569078947368421,
+      "grad_norm": 4.53125,
+      "grad_norm_var": 0.35949605305989585,
+      "learning_rate": 0.0001,
+      "loss": 3.0464,
+      "loss/crossentropy": 2.220982587337494,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22496577724814415,
+      "loss/reg": 0.0,
+      "step": 23850
+    },
+    {
+      "epoch": 0.15697368421052632,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.4110636393229167,
+      "learning_rate": 0.0001,
+      "loss": 3.0667,
+      "loss/crossentropy": 2.291455662250519,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2523698851466179,
+      "loss/reg": 0.0,
+      "step": 23860
+    },
+    {
+      "epoch": 0.15703947368421053,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.2355133056640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0867,
+      "loss/crossentropy": 2.4676188588142396,
+      "loss/hidden": 2.621875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20559777021408082,
+      "loss/reg": 0.0,
+      "step": 23870
+    },
+    {
+      "epoch": 0.15710526315789475,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.09799702962239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0123,
+      "loss/crossentropy": 2.4020154595375063,
+      "loss/hidden": 2.98125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24061929136514665,
+      "loss/reg": 0.0,
+      "step": 23880
+    },
+    {
+      "epoch": 0.15717105263157893,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 2.1426422119140627,
+      "learning_rate": 0.0001,
+      "loss": 3.02,
+      "loss/crossentropy": 2.0822024583816527,
+      "loss/hidden": 2.7421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23049053847789763,
+      "loss/reg": 0.0,
+      "step": 23890
+    },
+    {
+      "epoch": 0.15723684210526315,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 2.0829254150390626,
+      "learning_rate": 0.0001,
+      "loss": 3.0151,
+      "loss/crossentropy": 2.210465407371521,
+      "loss/hidden": 3.0328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24976756423711777,
+      "loss/reg": 0.0,
+      "step": 23900
+    },
+    {
+      "epoch": 0.15730263157894736,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.5316731770833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0642,
+      "loss/crossentropy": 2.1934141278266908,
+      "loss/hidden": 2.5921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18757687732577324,
+      "loss/reg": 0.0,
+      "step": 23910
+    },
+    {
+      "epoch": 0.15736842105263157,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.5506998697916666,
+      "learning_rate": 0.0001,
+      "loss": 3.1533,
+      "loss/crossentropy": 2.2942312955856323,
+      "loss/hidden": 3.0890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25652203559875486,
+      "loss/reg": 0.0,
+      "step": 23920
+    },
+    {
+      "epoch": 0.1574342105263158,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.38448893229166664,
+      "learning_rate": 0.0001,
+      "loss": 3.0496,
+      "loss/crossentropy": 2.4290681004524233,
+      "loss/hidden": 2.7046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22672214657068251,
+      "loss/reg": 0.0,
+      "step": 23930
+    },
+    {
+      "epoch": 0.1575,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.35227864583333335,
+      "learning_rate": 0.0001,
+      "loss": 3.0868,
+      "loss/crossentropy": 2.1520013570785523,
+      "loss/hidden": 2.8453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2738748073577881,
+      "loss/reg": 0.0,
+      "step": 23940
+    },
+    {
+      "epoch": 0.15756578947368421,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.07700907389322917,
+      "learning_rate": 0.0001,
+      "loss": 3.1436,
+      "loss/crossentropy": 2.5199429869651793,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2519889883697033,
+      "loss/reg": 0.0,
+      "step": 23950
+    },
+    {
+      "epoch": 0.15763157894736843,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 1.195580037434896,
+      "learning_rate": 0.0001,
+      "loss": 3.0924,
+      "loss/crossentropy": 2.362662875652313,
+      "loss/hidden": 2.946875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2809165805578232,
+      "loss/reg": 0.0,
+      "step": 23960
+    },
+    {
+      "epoch": 0.15769736842105264,
+      "grad_norm": 1.9921875,
+      "grad_norm_var": 1.2223609924316405,
+      "learning_rate": 0.0001,
+      "loss": 3.1196,
+      "loss/crossentropy": 2.360364031791687,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2587847799062729,
+      "loss/reg": 0.0,
+      "step": 23970
+    },
+    {
+      "epoch": 0.15776315789473686,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.028562164306640624,
+      "learning_rate": 0.0001,
+      "loss": 3.0014,
+      "loss/crossentropy": 2.268549180030823,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2278840996325016,
+      "loss/reg": 0.0,
+      "step": 23980
+    },
+    {
+      "epoch": 0.15782894736842104,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.14148661295572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0699,
+      "loss/crossentropy": 2.0497791051864622,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2620595782995224,
+      "loss/reg": 0.0,
+      "step": 23990
+    },
+    {
+      "epoch": 0.15789473684210525,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 2.927515672594874e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.2727,
+      "loss/crossentropy": 2.035860624909401,
+      "loss/hidden": 2.7390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20649760514497756,
+      "loss/reg": 0.0,
+      "step": 24000
+    },
+    {
+      "epoch": 0.15796052631578947,
+      "grad_norm": 3.265625,
+      "grad_norm_var": 2.927515672138351e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1828,
+      "loss/crossentropy": 2.358655941486359,
+      "loss/hidden": 2.8453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23838587403297423,
+      "loss/reg": 0.0,
+      "step": 24010
+    },
+    {
+      "epoch": 0.15802631578947368,
+      "grad_norm": 4.34375,
+      "grad_norm_var": 0.4474029541015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0939,
+      "loss/crossentropy": 2.436113882064819,
+      "loss/hidden": 2.734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25293364077806474,
+      "loss/reg": 0.0,
+      "step": 24020
+    },
+    {
+      "epoch": 0.1580921052631579,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.346875,
+      "learning_rate": 0.0001,
+      "loss": 3.1031,
+      "loss/crossentropy": 2.1950599789619445,
+      "loss/hidden": 3.025,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29097774922847747,
+      "loss/reg": 0.0,
+      "step": 24030
+    },
+    {
+      "epoch": 0.1581578947368421,
+      "grad_norm": 3.5,
+      "grad_norm_var": 0.84117431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.138,
+      "loss/crossentropy": 2.488498842716217,
+      "loss/hidden": 2.9921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2546138882637024,
+      "loss/reg": 0.0,
+      "step": 24040
+    },
+    {
+      "epoch": 0.15822368421052632,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.20335286458333332,
+      "learning_rate": 0.0001,
+      "loss": 3.1208,
+      "loss/crossentropy": 2.1634037137031554,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23214713484048843,
+      "loss/reg": 0.0,
+      "step": 24050
+    },
+    {
+      "epoch": 0.15828947368421054,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.036454264322916666,
+      "learning_rate": 0.0001,
+      "loss": 3.1708,
+      "loss/crossentropy": 2.3869518637657166,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2666103199124336,
+      "loss/reg": 0.0,
+      "step": 24060
+    },
+    {
+      "epoch": 0.15835526315789475,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.1547515869140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0924,
+      "loss/crossentropy": 2.2607127904891966,
+      "loss/hidden": 2.84375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25745444297790526,
+      "loss/reg": 0.0,
+      "step": 24070
+    },
+    {
+      "epoch": 0.15842105263157893,
+      "grad_norm": 4.0,
+      "grad_norm_var": 0.22906494140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0239,
+      "loss/crossentropy": 2.2656429171562196,
+      "loss/hidden": 2.7203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2332296848297119,
+      "loss/reg": 0.0,
+      "step": 24080
+    },
+    {
+      "epoch": 0.15848684210526315,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.23699442545572916,
+      "learning_rate": 0.0001,
+      "loss": 3.0181,
+      "loss/crossentropy": 2.2533790946006773,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2367706872522831,
+      "loss/reg": 0.0,
+      "step": 24090
+    },
+    {
+      "epoch": 0.15855263157894736,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.06172587076822917,
+      "learning_rate": 0.0001,
+      "loss": 3.045,
+      "loss/crossentropy": 2.3126723527908326,
+      "loss/hidden": 2.703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23582341223955156,
+      "loss/reg": 0.0,
+      "step": 24100
+    },
+    {
+      "epoch": 0.15861842105263158,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.08645426432291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0989,
+      "loss/crossentropy": 2.284153974056244,
+      "loss/hidden": 2.965625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3264383256435394,
+      "loss/reg": 0.0,
+      "step": 24110
+    },
+    {
+      "epoch": 0.1586842105263158,
+      "grad_norm": 3.671875,
+      "grad_norm_var": 172.1703125,
+      "learning_rate": 0.0001,
+      "loss": 3.2515,
+      "loss/crossentropy": 2.463814842700958,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2696057379245758,
+      "loss/reg": 0.0,
+      "step": 24120
+    },
+    {
+      "epoch": 0.15875,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.19868062337239584,
+      "learning_rate": 0.0001,
+      "loss": 3.1115,
+      "loss/crossentropy": 2.41875559091568,
+      "loss/hidden": 2.75625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2305321291089058,
+      "loss/reg": 0.0,
+      "step": 24130
+    },
+    {
+      "epoch": 0.15881578947368422,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.5208513895670573,
+      "learning_rate": 0.0001,
+      "loss": 3.0293,
+      "loss/crossentropy": 2.051196885108948,
+      "loss/hidden": 2.83125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22749217078089715,
+      "loss/reg": 0.0,
+      "step": 24140
+    },
+    {
+      "epoch": 0.15888157894736843,
+      "grad_norm": 2298478592.0,
+      "grad_norm_var": 3.301877386622895e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.3081,
+      "loss/crossentropy": 2.3360164284706117,
+      "loss/hidden": 3.290625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3266043797135353,
+      "loss/reg": 0.0,
+      "step": 24150
+    },
+    {
+      "epoch": 0.15894736842105264,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 3.3018773920608314e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0123,
+      "loss/crossentropy": 2.5075667977333067,
+      "loss/hidden": 2.603125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19614752382040024,
+      "loss/reg": 0.0,
+      "step": 24160
+    },
+    {
+      "epoch": 0.15901315789473683,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.06570002237955729,
+      "learning_rate": 0.0001,
+      "loss": 3.0729,
+      "loss/crossentropy": 2.1672366857528687,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2613453507423401,
+      "loss/reg": 0.0,
+      "step": 24170
+    },
+    {
+      "epoch": 0.15907894736842104,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.437475331624349,
+      "learning_rate": 0.0001,
+      "loss": 3.073,
+      "loss/crossentropy": 2.1420519262552262,
+      "loss/hidden": 2.7140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20748247653245927,
+      "loss/reg": 0.0,
+      "step": 24180
+    },
+    {
+      "epoch": 0.15914473684210526,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.5490468343098959,
+      "learning_rate": 0.0001,
+      "loss": 3.1018,
+      "loss/crossentropy": 2.5135774850845336,
+      "loss/hidden": 2.98125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2358848750591278,
+      "loss/reg": 0.0,
+      "step": 24190
+    },
+    {
+      "epoch": 0.15921052631578947,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.05199559529622396,
+      "learning_rate": 0.0001,
+      "loss": 3.0415,
+      "loss/crossentropy": 2.392611360549927,
+      "loss/hidden": 2.8,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29126308858394623,
+      "loss/reg": 0.0,
+      "step": 24200
+    },
+    {
+      "epoch": 0.15927631578947368,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.3502886454264323,
+      "learning_rate": 0.0001,
+      "loss": 3.0974,
+      "loss/crossentropy": 2.2008650302886963,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2575715593993664,
+      "loss/reg": 0.0,
+      "step": 24210
+    },
+    {
+      "epoch": 0.1593421052631579,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.2436920166015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1615,
+      "loss/crossentropy": 2.398923659324646,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2272189199924469,
+      "loss/reg": 0.0,
+      "step": 24220
+    },
+    {
+      "epoch": 0.1594078947368421,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.2711252848307292,
+      "learning_rate": 0.0001,
+      "loss": 3.1147,
+      "loss/crossentropy": 2.2159452080726623,
+      "loss/hidden": 2.90625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24379997029900552,
+      "loss/reg": 0.0,
+      "step": 24230
+    },
+    {
+      "epoch": 0.15947368421052632,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.16237691243489583,
+      "learning_rate": 0.0001,
+      "loss": 3.032,
+      "loss/crossentropy": 2.219667136669159,
+      "loss/hidden": 2.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2103893890976906,
+      "loss/reg": 0.0,
+      "step": 24240
+    },
+    {
+      "epoch": 0.15953947368421054,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.035920206705729166,
+      "learning_rate": 0.0001,
+      "loss": 3.1149,
+      "loss/crossentropy": 2.0510872304439545,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2106800675392151,
+      "loss/reg": 0.0,
+      "step": 24250
+    },
+    {
+      "epoch": 0.15960526315789475,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.05742085774739583,
+      "learning_rate": 0.0001,
+      "loss": 2.9921,
+      "loss/crossentropy": 2.202706849575043,
+      "loss/hidden": 2.6953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.200262913107872,
+      "loss/reg": 0.0,
+      "step": 24260
+    },
+    {
+      "epoch": 0.15967105263157894,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.47508138020833335,
+      "learning_rate": 0.0001,
+      "loss": 3.0355,
+      "loss/crossentropy": 2.312657380104065,
+      "loss/hidden": 2.95625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25428467951714995,
+      "loss/reg": 0.0,
+      "step": 24270
+    },
+    {
+      "epoch": 0.15973684210526315,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.11523030598958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1753,
+      "loss/crossentropy": 2.6283345460891723,
+      "loss/hidden": 2.9140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2865023031830788,
+      "loss/reg": 0.0,
+      "step": 24280
+    },
+    {
+      "epoch": 0.15980263157894736,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.04806315104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0594,
+      "loss/crossentropy": 2.4725473642349245,
+      "loss/hidden": 2.6453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20711587965488434,
+      "loss/reg": 0.0,
+      "step": 24290
+    },
+    {
+      "epoch": 0.15986842105263158,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 2.7487790628399786e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.265,
+      "loss/crossentropy": 2.2610708713531493,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27779700309038163,
+      "loss/reg": 0.0,
+      "step": 24300
+    },
+    {
+      "epoch": 0.1599342105263158,
+      "grad_norm": 1.828125,
+      "grad_norm_var": 2.748779062992896e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0342,
+      "loss/crossentropy": 2.4125430822372436,
+      "loss/hidden": 2.7078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24020388573408127,
+      "loss/reg": 0.0,
+      "step": 24310
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.10875244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0387,
+      "loss/crossentropy": 2.2118794441223146,
+      "loss/hidden": 2.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23726068288087845,
+      "loss/reg": 0.0,
+      "step": 24320
+    },
+    {
+      "epoch": 0.16006578947368422,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.052897135416666664,
+      "learning_rate": 0.0001,
+      "loss": 3.0208,
+      "loss/crossentropy": 2.4390548706054687,
+      "loss/hidden": 2.83125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24900538772344588,
+      "loss/reg": 0.0,
+      "step": 24330
+    },
+    {
+      "epoch": 0.16013157894736843,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.5782511393229167,
+      "learning_rate": 0.0001,
+      "loss": 3.2442,
+      "loss/crossentropy": 2.3729852437973022,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23976370990276336,
+      "loss/reg": 0.0,
+      "step": 24340
+    },
+    {
+      "epoch": 0.16019736842105264,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.11790364583333333,
+      "learning_rate": 0.0001,
+      "loss": 3.0854,
+      "loss/crossentropy": 2.4694011569023133,
+      "loss/hidden": 2.7875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25265379548072814,
+      "loss/reg": 0.0,
+      "step": 24350
+    },
+    {
+      "epoch": 0.16026315789473683,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.0570465087890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1662,
+      "loss/crossentropy": 2.044074738025665,
+      "loss/hidden": 2.959375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24245730862021447,
+      "loss/reg": 0.0,
+      "step": 24360
+    },
+    {
+      "epoch": 0.16032894736842104,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.0236724853515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0431,
+      "loss/crossentropy": 2.3142729878425596,
+      "loss/hidden": 2.7140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22269122749567033,
+      "loss/reg": 0.0,
+      "step": 24370
+    },
+    {
+      "epoch": 0.16039473684210526,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.08518473307291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1019,
+      "loss/crossentropy": 2.0431338131427763,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2301457904279232,
+      "loss/reg": 0.0,
+      "step": 24380
+    },
+    {
+      "epoch": 0.16046052631578947,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.08948160807291666,
+      "learning_rate": 0.0001,
+      "loss": 2.9974,
+      "loss/crossentropy": 2.215762954950333,
+      "loss/hidden": 2.6671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21034832447767257,
+      "loss/reg": 0.0,
+      "step": 24390
+    },
+    {
+      "epoch": 0.16052631578947368,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.013671875,
+      "learning_rate": 0.0001,
+      "loss": 3.0257,
+      "loss/crossentropy": 2.285198521614075,
+      "loss/hidden": 2.7625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22049440741539,
+      "loss/reg": 0.0,
+      "step": 24400
+    },
+    {
+      "epoch": 0.1605921052631579,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.04234110514322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0891,
+      "loss/crossentropy": 2.509168195724487,
+      "loss/hidden": 2.95,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28251273930072784,
+      "loss/reg": 0.0,
+      "step": 24410
+    },
+    {
+      "epoch": 0.1606578947368421,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.33464253743489586,
+      "learning_rate": 0.0001,
+      "loss": 3.0359,
+      "loss/crossentropy": 2.0764550805091857,
+      "loss/hidden": 2.93125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2346973106265068,
+      "loss/reg": 0.0,
+      "step": 24420
+    },
+    {
+      "epoch": 0.16072368421052632,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.04140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0962,
+      "loss/crossentropy": 2.4705166459083556,
+      "loss/hidden": 2.8,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2445474848151207,
+      "loss/reg": 0.0,
+      "step": 24430
+    },
+    {
+      "epoch": 0.16078947368421054,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.3681060791015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0701,
+      "loss/crossentropy": 2.347282111644745,
+      "loss/hidden": 2.9,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27417414337396623,
+      "loss/reg": 0.0,
+      "step": 24440
+    },
+    {
+      "epoch": 0.16085526315789472,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.0697662353515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1662,
+      "loss/crossentropy": 2.1282618045806885,
+      "loss/hidden": 3.0671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32502798140048983,
+      "loss/reg": 0.0,
+      "step": 24450
+    },
+    {
+      "epoch": 0.16092105263157894,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.06052958170572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0394,
+      "loss/crossentropy": 2.5319641828536987,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23511321395635604,
+      "loss/reg": 0.0,
+      "step": 24460
+    },
+    {
+      "epoch": 0.16098684210526315,
+      "grad_norm": 3.84375,
+      "grad_norm_var": 0.18154067993164064,
+      "learning_rate": 0.0001,
+      "loss": 3.0401,
+      "loss/crossentropy": 2.3912153840065002,
+      "loss/hidden": 2.828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2340209573507309,
+      "loss/reg": 0.0,
+      "step": 24470
+    },
+    {
+      "epoch": 0.16105263157894736,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.45718485514322915,
+      "learning_rate": 0.0001,
+      "loss": 3.0803,
+      "loss/crossentropy": 2.4838199734687807,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23594623059034348,
+      "loss/reg": 0.0,
+      "step": 24480
+    },
+    {
+      "epoch": 0.16111842105263158,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.12635269165039062,
+      "learning_rate": 0.0001,
+      "loss": 3.0486,
+      "loss/crossentropy": 2.2548271775245667,
+      "loss/hidden": 2.8234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22111207991838455,
+      "loss/reg": 0.0,
+      "step": 24490
+    },
+    {
+      "epoch": 0.1611842105263158,
+      "grad_norm": 1.953125,
+      "grad_norm_var": 0.08367487589518229,
+      "learning_rate": 0.0001,
+      "loss": 3.0843,
+      "loss/crossentropy": 2.364280033111572,
+      "loss/hidden": 2.6921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2182385429739952,
+      "loss/reg": 0.0,
+      "step": 24500
+    },
+    {
+      "epoch": 0.16125,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.13824462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1481,
+      "loss/crossentropy": 2.3593606472015383,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24336900562047958,
+      "loss/reg": 0.0,
+      "step": 24510
+    },
+    {
+      "epoch": 0.16131578947368422,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.24010009765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0484,
+      "loss/crossentropy": 2.212929093837738,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23406590819358825,
+      "loss/reg": 0.0,
+      "step": 24520
+    },
+    {
+      "epoch": 0.16138157894736843,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.240869140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1152,
+      "loss/crossentropy": 2.0746700882911684,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20809975266456604,
+      "loss/reg": 0.0,
+      "step": 24530
+    },
+    {
+      "epoch": 0.16144736842105264,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.1971343994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0887,
+      "loss/crossentropy": 2.5128459453582765,
+      "loss/hidden": 2.7875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23968077450990677,
+      "loss/reg": 0.0,
+      "step": 24540
+    },
+    {
+      "epoch": 0.16151315789473683,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.10077718098958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0776,
+      "loss/crossentropy": 2.2077295780181885,
+      "loss/hidden": 2.91875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24557382240891457,
+      "loss/reg": 0.0,
+      "step": 24550
+    },
+    {
+      "epoch": 0.16157894736842104,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.09933268229166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0416,
+      "loss/crossentropy": 2.3417739272117615,
+      "loss/hidden": 2.715625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23852093145251274,
+      "loss/reg": 0.0,
+      "step": 24560
+    },
+    {
+      "epoch": 0.16164473684210526,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.16413472493489584,
+      "learning_rate": 0.0001,
+      "loss": 3.1237,
+      "loss/crossentropy": 2.230830729007721,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23145168870687485,
+      "loss/reg": 0.0,
+      "step": 24570
+    },
+    {
+      "epoch": 0.16171052631578947,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.11370035807291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1152,
+      "loss/crossentropy": 2.1511558890342712,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24633188247680665,
+      "loss/reg": 0.0,
+      "step": 24580
+    },
+    {
+      "epoch": 0.16177631578947368,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.3695818583170573,
+      "learning_rate": 0.0001,
+      "loss": 3.0297,
+      "loss/crossentropy": 2.3748416185379027,
+      "loss/hidden": 2.7984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2599245056509972,
+      "loss/reg": 0.0,
+      "step": 24590
+    },
+    {
+      "epoch": 0.1618421052631579,
+      "grad_norm": 1.96875,
+      "grad_norm_var": 0.11649576822916667,
+      "learning_rate": 0.0001,
+      "loss": 3.0334,
+      "loss/crossentropy": 1.958118262887001,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2173558861017227,
+      "loss/reg": 0.0,
+      "step": 24600
+    },
+    {
+      "epoch": 0.1619078947368421,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.09065729777018229,
+      "learning_rate": 0.0001,
+      "loss": 3.0806,
+      "loss/crossentropy": 2.3804409861564637,
+      "loss/hidden": 2.84375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28529137223958967,
+      "loss/reg": 0.0,
+      "step": 24610
+    },
+    {
+      "epoch": 0.16197368421052633,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.18765869140625,
+      "learning_rate": 0.0001,
+      "loss": 3.1415,
+      "loss/crossentropy": 2.4188234567642213,
+      "loss/hidden": 2.840625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2486872687935829,
+      "loss/reg": 0.0,
+      "step": 24620
+    },
+    {
+      "epoch": 0.16203947368421054,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.1975982666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0741,
+      "loss/crossentropy": 2.2737186551094055,
+      "loss/hidden": 2.7578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21121124178171158,
+      "loss/reg": 0.0,
+      "step": 24630
+    },
+    {
+      "epoch": 0.16210526315789472,
+      "grad_norm": 2.5,
+      "grad_norm_var": 2.4922159830729167,
+      "learning_rate": 0.0001,
+      "loss": 3.0672,
+      "loss/crossentropy": 1.935386747121811,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21296948455274106,
+      "loss/reg": 0.0,
+      "step": 24640
+    },
+    {
+      "epoch": 0.16217105263157894,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.053587849934895834,
+      "learning_rate": 0.0001,
+      "loss": 3.1024,
+      "loss/crossentropy": 2.456861126422882,
+      "loss/hidden": 2.9140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29705640524625776,
+      "loss/reg": 0.0,
+      "step": 24650
+    },
+    {
+      "epoch": 0.16223684210526315,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.2731679280598958,
+      "learning_rate": 0.0001,
+      "loss": 3.0683,
+      "loss/crossentropy": 2.3239927887916565,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24205588102340697,
+      "loss/reg": 0.0,
+      "step": 24660
+    },
+    {
+      "epoch": 0.16230263157894737,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.22311197916666667,
+      "learning_rate": 0.0001,
+      "loss": 3.0314,
+      "loss/crossentropy": 2.3312941789627075,
+      "loss/hidden": 2.971875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28574763536453246,
+      "loss/reg": 0.0,
+      "step": 24670
+    },
+    {
+      "epoch": 0.16236842105263158,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.24403889973958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1553,
+      "loss/crossentropy": 2.493465173244476,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23360646292567253,
+      "loss/reg": 0.0,
+      "step": 24680
+    },
+    {
+      "epoch": 0.1624342105263158,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.16258138020833332,
+      "learning_rate": 0.0001,
+      "loss": 3.088,
+      "loss/crossentropy": 2.3201822876930236,
+      "loss/hidden": 2.840625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24165551662445067,
+      "loss/reg": 0.0,
+      "step": 24690
+    },
+    {
+      "epoch": 0.1625,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.18403218587239584,
+      "learning_rate": 0.0001,
+      "loss": 3.1727,
+      "loss/crossentropy": 2.2391496330499647,
+      "loss/hidden": 2.8390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25802505761384964,
+      "loss/reg": 0.0,
+      "step": 24700
+    },
+    {
+      "epoch": 0.16256578947368422,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.047663370768229164,
+      "learning_rate": 0.0001,
+      "loss": 3.0701,
+      "loss/crossentropy": 2.1174232959747314,
+      "loss/hidden": 2.8875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25788239687681197,
+      "loss/reg": 0.0,
+      "step": 24710
+    },
+    {
+      "epoch": 0.16263157894736843,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.03251520792643229,
+      "learning_rate": 0.0001,
+      "loss": 3.0028,
+      "loss/crossentropy": 2.303908097743988,
+      "loss/hidden": 2.6953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22646306753158568,
+      "loss/reg": 0.0,
+      "step": 24720
+    },
+    {
+      "epoch": 0.16269736842105262,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.09428609212239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0573,
+      "loss/crossentropy": 2.4429296493530273,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24618532359600068,
+      "loss/reg": 0.0,
+      "step": 24730
+    },
+    {
+      "epoch": 0.16276315789473683,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.10568745930989583,
+      "learning_rate": 0.0001,
+      "loss": 3.0411,
+      "loss/crossentropy": 2.245293366909027,
+      "loss/hidden": 3.0265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3406794846057892,
+      "loss/reg": 0.0,
+      "step": 24740
+    },
+    {
+      "epoch": 0.16282894736842105,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.24233169555664064,
+      "learning_rate": 0.0001,
+      "loss": 3.0254,
+      "loss/crossentropy": 2.1815924525260924,
+      "loss/hidden": 2.7078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20982651114463807,
+      "loss/reg": 0.0,
+      "step": 24750
+    },
+    {
+      "epoch": 0.16289473684210526,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.051889801025390626,
+      "learning_rate": 0.0001,
+      "loss": 3.0337,
+      "loss/crossentropy": 2.530372714996338,
+      "loss/hidden": 2.6765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22411757558584214,
+      "loss/reg": 0.0,
+      "step": 24760
+    },
+    {
+      "epoch": 0.16296052631578947,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.025414021809895833,
+      "learning_rate": 0.0001,
+      "loss": 2.9668,
+      "loss/crossentropy": 2.271071231365204,
+      "loss/hidden": 2.8390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23753461390733718,
+      "loss/reg": 0.0,
+      "step": 24770
+    },
+    {
+      "epoch": 0.16302631578947369,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.0559722900390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0646,
+      "loss/crossentropy": 2.2592864990234376,
+      "loss/hidden": 2.83125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21606265306472777,
+      "loss/reg": 0.0,
+      "step": 24780
+    },
+    {
+      "epoch": 0.1630921052631579,
+      "grad_norm": 1.9921875,
+      "grad_norm_var": 0.05602595011393229,
+      "learning_rate": 0.0001,
+      "loss": 3.059,
+      "loss/crossentropy": 2.021050810813904,
+      "loss/hidden": 2.8390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2471683457493782,
+      "loss/reg": 0.0,
+      "step": 24790
+    },
+    {
+      "epoch": 0.1631578947368421,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.25566991170247394,
+      "learning_rate": 0.0001,
+      "loss": 3.0148,
+      "loss/crossentropy": 2.4111449480056764,
+      "loss/hidden": 2.678125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23949681222438812,
+      "loss/reg": 0.0,
+      "step": 24800
+    },
+    {
+      "epoch": 0.16322368421052633,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.21315816243489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0918,
+      "loss/crossentropy": 2.5244083642959594,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2454654648900032,
+      "loss/reg": 0.0,
+      "step": 24810
+    },
+    {
+      "epoch": 0.16328947368421054,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.09656575520833334,
+      "learning_rate": 0.0001,
+      "loss": 2.9972,
+      "loss/crossentropy": 2.4637012243270875,
+      "loss/hidden": 3.0046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2438522458076477,
+      "loss/reg": 0.0,
+      "step": 24820
+    },
+    {
+      "epoch": 0.16335526315789473,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.03803609212239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0278,
+      "loss/crossentropy": 2.2508904099464417,
+      "loss/hidden": 2.775,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24632504507899283,
+      "loss/reg": 0.0,
+      "step": 24830
+    },
+    {
+      "epoch": 0.16342105263157894,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.0544921875,
+      "learning_rate": 0.0001,
+      "loss": 3.0273,
+      "loss/crossentropy": 2.2554102897644044,
+      "loss/hidden": 2.696875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2244628369808197,
+      "loss/reg": 0.0,
+      "step": 24840
+    },
+    {
+      "epoch": 0.16348684210526315,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.0915679931640625,
+      "learning_rate": 0.0001,
+      "loss": 2.9974,
+      "loss/crossentropy": 2.4115317344665526,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22097424566745758,
+      "loss/reg": 0.0,
+      "step": 24850
+    },
+    {
+      "epoch": 0.16355263157894737,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.0794342041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0875,
+      "loss/crossentropy": 2.2210743844509127,
+      "loss/hidden": 2.9234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2252424478530884,
+      "loss/reg": 0.0,
+      "step": 24860
+    },
+    {
+      "epoch": 0.16361842105263158,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 3.853392423706864e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1398,
+      "loss/crossentropy": 2.5947876930236817,
+      "loss/hidden": 3.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33976440876722336,
+      "loss/reg": 0.0,
+      "step": 24870
+    },
+    {
+      "epoch": 0.1636842105263158,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 3.853392423689082e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0576,
+      "loss/crossentropy": 2.177506458759308,
+      "loss/hidden": 2.8671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.254705648124218,
+      "loss/reg": 0.0,
+      "step": 24880
+    },
+    {
+      "epoch": 0.16375,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.08782730102539063,
+      "learning_rate": 0.0001,
+      "loss": 3.0306,
+      "loss/crossentropy": 2.4746748208999634,
+      "loss/hidden": 2.78125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2408411219716072,
+      "loss/reg": 0.0,
+      "step": 24890
+    },
+    {
+      "epoch": 0.16381578947368422,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.07857666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0083,
+      "loss/crossentropy": 2.273878073692322,
+      "loss/hidden": 2.9171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2608170732855797,
+      "loss/reg": 0.0,
+      "step": 24900
+    },
+    {
+      "epoch": 0.16388157894736843,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.17737528483072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1428,
+      "loss/crossentropy": 2.240007960796356,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22373474836349488,
+      "loss/reg": 0.0,
+      "step": 24910
+    },
+    {
+      "epoch": 0.16394736842105262,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.10565999348958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0246,
+      "loss/crossentropy": 2.229018306732178,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29760813266038894,
+      "loss/reg": 0.0,
+      "step": 24920
+    },
+    {
+      "epoch": 0.16401315789473683,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.3898834228515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1527,
+      "loss/crossentropy": 2.2012178540229796,
+      "loss/hidden": 2.959375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28048097975552083,
+      "loss/reg": 0.0,
+      "step": 24930
+    },
+    {
+      "epoch": 0.16407894736842105,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.1896636962890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1143,
+      "loss/crossentropy": 2.396012032032013,
+      "loss/hidden": 2.984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31100255995988846,
+      "loss/reg": 0.0,
+      "step": 24940
+    },
+    {
+      "epoch": 0.16414473684210526,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.11585464477539062,
+      "learning_rate": 0.0001,
+      "loss": 2.9923,
+      "loss/crossentropy": 2.35096116065979,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1970454752445221,
+      "loss/reg": 0.0,
+      "step": 24950
+    },
+    {
+      "epoch": 0.16421052631578947,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.048278554280598955,
+      "learning_rate": 0.0001,
+      "loss": 3.0105,
+      "loss/crossentropy": 2.1574088990688325,
+      "loss/hidden": 2.8,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20654048770666122,
+      "loss/reg": 0.0,
+      "step": 24960
+    },
+    {
+      "epoch": 0.1642763157894737,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.41565348307291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1106,
+      "loss/crossentropy": 2.087399756908417,
+      "loss/hidden": 3.11875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2551331013441086,
+      "loss/reg": 0.0,
+      "step": 24970
+    },
+    {
+      "epoch": 0.1643421052631579,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.4007232666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1262,
+      "loss/crossentropy": 2.104507529735565,
+      "loss/hidden": 3.0828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3040700241923332,
+      "loss/reg": 0.0,
+      "step": 24980
+    },
+    {
+      "epoch": 0.16440789473684211,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.29108072916666666,
+      "learning_rate": 0.0001,
+      "loss": 3.1264,
+      "loss/crossentropy": 2.393839418888092,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2309044197201729,
+      "loss/reg": 0.0,
+      "step": 24990
+    },
+    {
+      "epoch": 0.16447368421052633,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.34716695149739585,
+      "learning_rate": 0.0001,
+      "loss": 3.0873,
+      "loss/crossentropy": 2.295112156867981,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2622707739472389,
+      "loss/reg": 0.0,
+      "step": 25000
+    },
+    {
+      "epoch": 0.1645394736842105,
+      "grad_norm": 2.9375,
+      "grad_norm_var": 2.9363606770833335,
+      "learning_rate": 0.0001,
+      "loss": 3.1494,
+      "loss/crossentropy": 2.0088412761688232,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2086470142006874,
+      "loss/reg": 0.0,
+      "step": 25010
+    },
+    {
+      "epoch": 0.16460526315789473,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 3.0145467122395835,
+      "learning_rate": 0.0001,
+      "loss": 3.0267,
+      "loss/crossentropy": 2.230376589298248,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25718885809183123,
+      "loss/reg": 0.0,
+      "step": 25020
+    },
+    {
+      "epoch": 0.16467105263157894,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.16364644368489584,
+      "learning_rate": 0.0001,
+      "loss": 3.1098,
+      "loss/crossentropy": 2.2357122182846068,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20328374579548836,
+      "loss/reg": 0.0,
+      "step": 25030
+    },
+    {
+      "epoch": 0.16473684210526315,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.13963597615559895,
+      "learning_rate": 0.0001,
+      "loss": 3.0687,
+      "loss/crossentropy": 2.3226524710655214,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23719747960567475,
+      "loss/reg": 0.0,
+      "step": 25040
+    },
+    {
+      "epoch": 0.16480263157894737,
+      "grad_norm": 2298478592.0,
+      "grad_norm_var": 3.301877391914183e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.151,
+      "loss/crossentropy": 2.2604412317276,
+      "loss/hidden": 2.9375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2530272454023361,
+      "loss/reg": 0.0,
+      "step": 25050
+    },
+    {
+      "epoch": 0.16486842105263158,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 3.301877391448801e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0688,
+      "loss/crossentropy": 2.374041938781738,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23906515687704086,
+      "loss/reg": 0.0,
+      "step": 25060
+    },
+    {
+      "epoch": 0.1649342105263158,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.08870035807291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0466,
+      "loss/crossentropy": 2.2033262491226195,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2643999807536602,
+      "loss/reg": 0.0,
+      "step": 25070
+    },
+    {
+      "epoch": 0.165,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.22410481770833332,
+      "learning_rate": 0.0001,
+      "loss": 3.0549,
+      "loss/crossentropy": 2.268805372714996,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22531848698854445,
+      "loss/reg": 0.0,
+      "step": 25080
+    },
+    {
+      "epoch": 0.16506578947368422,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.17864481608072916,
+      "learning_rate": 0.0001,
+      "loss": 3.05,
+      "loss/crossentropy": 2.32329341173172,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23195795714855194,
+      "loss/reg": 0.0,
+      "step": 25090
+    },
+    {
+      "epoch": 0.16513157894736843,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.07084859212239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0737,
+      "loss/crossentropy": 2.3633965373039247,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24620314687490463,
+      "loss/reg": 0.0,
+      "step": 25100
+    },
+    {
+      "epoch": 0.16519736842105262,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.37231343587239585,
+      "learning_rate": 0.0001,
+      "loss": 3.0891,
+      "loss/crossentropy": 2.4329662203788756,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23993222489953042,
+      "loss/reg": 0.0,
+      "step": 25110
+    },
+    {
+      "epoch": 0.16526315789473683,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.48580322265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0705,
+      "loss/crossentropy": 2.418682646751404,
+      "loss/hidden": 2.753125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22389767169952393,
+      "loss/reg": 0.0,
+      "step": 25120
+    },
+    {
+      "epoch": 0.16532894736842105,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.08493550618489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0513,
+      "loss/crossentropy": 2.394873285293579,
+      "loss/hidden": 2.828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2614659383893013,
+      "loss/reg": 0.0,
+      "step": 25130
+    },
+    {
+      "epoch": 0.16539473684210526,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.19609273274739583,
+      "learning_rate": 0.0001,
+      "loss": 3.0834,
+      "loss/crossentropy": 2.1957722663879395,
+      "loss/hidden": 2.9609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2501178741455078,
+      "loss/reg": 0.0,
+      "step": 25140
+    },
+    {
+      "epoch": 0.16546052631578947,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.17849019368489583,
+      "learning_rate": 0.0001,
+      "loss": 3.024,
+      "loss/crossentropy": 2.389811897277832,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2392006903886795,
+      "loss/reg": 0.0,
+      "step": 25150
+    },
+    {
+      "epoch": 0.1655263157894737,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.08551610310872396,
+      "learning_rate": 0.0001,
+      "loss": 3.0152,
+      "loss/crossentropy": 2.328965938091278,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2474894031882286,
+      "loss/reg": 0.0,
+      "step": 25160
+    },
+    {
+      "epoch": 0.1655921052631579,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.26106363932291665,
+      "learning_rate": 0.0001,
+      "loss": 3.092,
+      "loss/crossentropy": 2.5120302557945253,
+      "loss/hidden": 3.0421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25525424629449844,
+      "loss/reg": 0.0,
+      "step": 25170
+    },
+    {
+      "epoch": 0.16565789473684212,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.4998372395833333,
+      "learning_rate": 0.0001,
+      "loss": 3.1112,
+      "loss/crossentropy": 2.37037136554718,
+      "loss/hidden": 2.6546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22218506336212157,
+      "loss/reg": 0.0,
+      "step": 25180
+    },
+    {
+      "epoch": 0.16572368421052633,
+      "grad_norm": 4.6875,
+      "grad_norm_var": 0.4266916910807292,
+      "learning_rate": 0.0001,
+      "loss": 3.1019,
+      "loss/crossentropy": 2.3847400188446044,
+      "loss/hidden": 3.021875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2786879613995552,
+      "loss/reg": 0.0,
+      "step": 25190
+    },
+    {
+      "epoch": 0.16578947368421051,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.4002766927083333,
+      "learning_rate": 0.0001,
+      "loss": 3.0387,
+      "loss/crossentropy": 2.2348836183547975,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24782560914754867,
+      "loss/reg": 0.0,
+      "step": 25200
+    },
+    {
+      "epoch": 0.16585526315789473,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.21065165201822916,
+      "learning_rate": 0.0001,
+      "loss": 3.1195,
+      "loss/crossentropy": 2.257748210430145,
+      "loss/hidden": 2.8140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21636683791875838,
+      "loss/reg": 0.0,
+      "step": 25210
+    },
+    {
+      "epoch": 0.16592105263157894,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.05474853515625,
+      "learning_rate": 0.0001,
+      "loss": 3.099,
+      "loss/crossentropy": 2.4620559096336363,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21316515058279037,
+      "loss/reg": 0.0,
+      "step": 25220
+    },
+    {
+      "epoch": 0.16598684210526315,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.03912353515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0173,
+      "loss/crossentropy": 2.3012343645095825,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2387869328260422,
+      "loss/reg": 0.0,
+      "step": 25230
+    },
+    {
+      "epoch": 0.16605263157894737,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.06116434733072917,
+      "learning_rate": 0.0001,
+      "loss": 3.0454,
+      "loss/crossentropy": 2.20231409072876,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22614135295152665,
+      "loss/reg": 0.0,
+      "step": 25240
+    },
+    {
+      "epoch": 0.16611842105263158,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.05122782389322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0264,
+      "loss/crossentropy": 2.2305002450942992,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22803680896759032,
+      "loss/reg": 0.0,
+      "step": 25250
+    },
+    {
+      "epoch": 0.1661842105263158,
+      "grad_norm": 2.0,
+      "grad_norm_var": 0.05510660807291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0045,
+      "loss/crossentropy": 2.209262716770172,
+      "loss/hidden": 2.871875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2319029837846756,
+      "loss/reg": 0.0,
+      "step": 25260
+    },
+    {
+      "epoch": 0.16625,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.0671051025390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0665,
+      "loss/crossentropy": 2.3621141076087953,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24575576186180115,
+      "loss/reg": 0.0,
+      "step": 25270
+    },
+    {
+      "epoch": 0.16631578947368422,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.1259674072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0746,
+      "loss/crossentropy": 2.297244334220886,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21904474049806594,
+      "loss/reg": 0.0,
+      "step": 25280
+    },
+    {
+      "epoch": 0.1663815789473684,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.12108968098958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0591,
+      "loss/crossentropy": 2.167328989505768,
+      "loss/hidden": 2.8234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.254831300675869,
+      "loss/reg": 0.0,
+      "step": 25290
+    },
+    {
+      "epoch": 0.16644736842105262,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.030272420247395834,
+      "learning_rate": 0.0001,
+      "loss": 3.0706,
+      "loss/crossentropy": 2.5636652946472167,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26798896938562394,
+      "loss/reg": 0.0,
+      "step": 25300
+    },
+    {
+      "epoch": 0.16651315789473684,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.03521728515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0397,
+      "loss/crossentropy": 2.241175901889801,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23810605853796005,
+      "loss/reg": 0.0,
+      "step": 25310
+    },
+    {
+      "epoch": 0.16657894736842105,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.1312652587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0874,
+      "loss/crossentropy": 2.461783027648926,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22298123091459274,
+      "loss/reg": 0.0,
+      "step": 25320
+    },
+    {
+      "epoch": 0.16664473684210526,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.01871337890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0339,
+      "loss/crossentropy": 2.6030556201934814,
+      "loss/hidden": 2.665625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2566943824291229,
+      "loss/reg": 0.0,
+      "step": 25330
+    },
+    {
+      "epoch": 0.16671052631578948,
+      "grad_norm": 5.375,
+      "grad_norm_var": 0.6346506754557292,
+      "learning_rate": 0.0001,
+      "loss": 3.0403,
+      "loss/crossentropy": 2.2260714411735534,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21501264423131944,
+      "loss/reg": 0.0,
+      "step": 25340
+    },
+    {
+      "epoch": 0.1667763157894737,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.6130198160807292,
+      "learning_rate": 0.0001,
+      "loss": 3.1302,
+      "loss/crossentropy": 2.2205970883369446,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2506598949432373,
+      "loss/reg": 0.0,
+      "step": 25350
+    },
+    {
+      "epoch": 0.1668421052631579,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.07870686848958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0579,
+      "loss/crossentropy": 2.3078066170215608,
+      "loss/hidden": 2.8890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2546592831611633,
+      "loss/reg": 0.0,
+      "step": 25360
+    },
+    {
+      "epoch": 0.16690789473684212,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.7152821858723958,
+      "learning_rate": 0.0001,
+      "loss": 3.084,
+      "loss/crossentropy": 2.103282463550568,
+      "loss/hidden": 3.0796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2719234719872475,
+      "loss/reg": 0.0,
+      "step": 25370
+    },
+    {
+      "epoch": 0.1669736842105263,
+      "grad_norm": 1.8984375,
+      "grad_norm_var": 1.6605242411295573,
+      "learning_rate": 0.0001,
+      "loss": 3.0337,
+      "loss/crossentropy": 2.533907437324524,
+      "loss/hidden": 2.6671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2172473669052124,
+      "loss/reg": 0.0,
+      "step": 25380
+    },
+    {
+      "epoch": 0.16703947368421052,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.841961415608724,
+      "learning_rate": 0.0001,
+      "loss": 3.1811,
+      "loss/crossentropy": 2.349000704288483,
+      "loss/hidden": 2.84375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2332908734679222,
+      "loss/reg": 0.0,
+      "step": 25390
+    },
+    {
+      "epoch": 0.16710526315789473,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.75146484375,
+      "learning_rate": 0.0001,
+      "loss": 3.0519,
+      "loss/crossentropy": 2.3112148702144624,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2460096523165703,
+      "loss/reg": 0.0,
+      "step": 25400
+    },
+    {
+      "epoch": 0.16717105263157894,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.0905426025390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0931,
+      "loss/crossentropy": 2.208834648132324,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21595955416560172,
+      "loss/reg": 0.0,
+      "step": 25410
+    },
+    {
+      "epoch": 0.16723684210526316,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.07610575358072917,
+      "learning_rate": 0.0001,
+      "loss": 3.0541,
+      "loss/crossentropy": 2.2315654158592224,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23381351083517074,
+      "loss/reg": 0.0,
+      "step": 25420
+    },
+    {
+      "epoch": 0.16730263157894737,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.043302154541015624,
+      "learning_rate": 0.0001,
+      "loss": 3.0197,
+      "loss/crossentropy": 2.3990158438682556,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24045146703720094,
+      "loss/reg": 0.0,
+      "step": 25430
+    },
+    {
+      "epoch": 0.16736842105263158,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.0761138916015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0792,
+      "loss/crossentropy": 2.1985233664512633,
+      "loss/hidden": 2.8265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2625061020255089,
+      "loss/reg": 0.0,
+      "step": 25440
+    },
+    {
+      "epoch": 0.1674342105263158,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.08713150024414062,
+      "learning_rate": 0.0001,
+      "loss": 3.0474,
+      "loss/crossentropy": 2.2765584170818327,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2559843085706234,
+      "loss/reg": 0.0,
+      "step": 25450
+    },
+    {
+      "epoch": 0.1675,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.06063003540039062,
+      "learning_rate": 0.0001,
+      "loss": 3.0819,
+      "loss/crossentropy": 2.241065299510956,
+      "loss/hidden": 2.815625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22642052918672562,
+      "loss/reg": 0.0,
+      "step": 25460
+    },
+    {
+      "epoch": 0.16756578947368422,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.5740468343098958,
+      "learning_rate": 0.0001,
+      "loss": 3.1251,
+      "loss/crossentropy": 2.388518822193146,
+      "loss/hidden": 3.06875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.283678263425827,
+      "loss/reg": 0.0,
+      "step": 25470
+    },
+    {
+      "epoch": 0.1676315789473684,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.13244400024414063,
+      "learning_rate": 0.0001,
+      "loss": 3.09,
+      "loss/crossentropy": 2.2575212955474853,
+      "loss/hidden": 2.790625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21582936197519303,
+      "loss/reg": 0.0,
+      "step": 25480
+    },
+    {
+      "epoch": 0.16769736842105262,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.14921773274739583,
+      "learning_rate": 0.0001,
+      "loss": 3.0472,
+      "loss/crossentropy": 2.378064513206482,
+      "loss/hidden": 2.70625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22392996102571489,
+      "loss/reg": 0.0,
+      "step": 25490
+    },
+    {
+      "epoch": 0.16776315789473684,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.046083323160807294,
+      "learning_rate": 0.0001,
+      "loss": 3.1247,
+      "loss/crossentropy": 2.354731285572052,
+      "loss/hidden": 2.8359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24077038019895552,
+      "loss/reg": 0.0,
+      "step": 25500
+    },
+    {
+      "epoch": 0.16782894736842105,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.024348958333333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0441,
+      "loss/crossentropy": 2.645008158683777,
+      "loss/hidden": 2.8,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2468577191233635,
+      "loss/reg": 0.0,
+      "step": 25510
+    },
+    {
+      "epoch": 0.16789473684210526,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.10139567057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.158,
+      "loss/crossentropy": 2.4286780834197996,
+      "loss/hidden": 3.015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33578050583601,
+      "loss/reg": 0.0,
+      "step": 25520
+    },
+    {
+      "epoch": 0.16796052631578948,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.03992513020833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0624,
+      "loss/crossentropy": 2.4203076124191285,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27344991117715833,
+      "loss/reg": 0.0,
+      "step": 25530
+    },
+    {
+      "epoch": 0.1680263157894737,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.03359273274739583,
+      "learning_rate": 0.0001,
+      "loss": 3.111,
+      "loss/crossentropy": 2.1932618856430053,
+      "loss/hidden": 3.00625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27270109951496124,
+      "loss/reg": 0.0,
+      "step": 25540
+    },
+    {
+      "epoch": 0.1680921052631579,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.03035252888997396,
+      "learning_rate": 0.0001,
+      "loss": 3.0572,
+      "loss/crossentropy": 2.1443075835704803,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21711303144693375,
+      "loss/reg": 0.0,
+      "step": 25550
+    },
+    {
+      "epoch": 0.16815789473684212,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.06535822550455729,
+      "learning_rate": 0.0001,
+      "loss": 3.0248,
+      "loss/crossentropy": 2.2873115301132203,
+      "loss/hidden": 2.940625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26505507379770277,
+      "loss/reg": 0.0,
+      "step": 25560
+    },
+    {
+      "epoch": 0.1682236842105263,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.21750895182291666,
+      "learning_rate": 0.0001,
+      "loss": 3.245,
+      "loss/crossentropy": 2.464116406440735,
+      "loss/hidden": 2.765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2560413718223572,
+      "loss/reg": 0.0,
+      "step": 25570
+    },
+    {
+      "epoch": 0.16828947368421052,
+      "grad_norm": 4.0625,
+      "grad_norm_var": 0.2865193684895833,
+      "learning_rate": 0.0001,
+      "loss": 3.1803,
+      "loss/crossentropy": 2.299769949913025,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2611501321196556,
+      "loss/reg": 0.0,
+      "step": 25580
+    },
+    {
+      "epoch": 0.16835526315789473,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.24952799479166668,
+      "learning_rate": 0.0001,
+      "loss": 3.1427,
+      "loss/crossentropy": 2.0992822468280794,
+      "loss/hidden": 3.109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2576856903731823,
+      "loss/reg": 0.0,
+      "step": 25590
+    },
+    {
+      "epoch": 0.16842105263157894,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.4341705322265625,
+      "learning_rate": 0.0001,
+      "loss": 3.101,
+      "loss/crossentropy": 2.0800092458724975,
+      "loss/hidden": 2.85625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20497069805860518,
+      "loss/reg": 0.0,
+      "step": 25600
+    },
+    {
+      "epoch": 0.16848684210526316,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.4129231770833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0582,
+      "loss/crossentropy": 2.3167444467544556,
+      "loss/hidden": 2.8640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2798536166548729,
+      "loss/reg": 0.0,
+      "step": 25610
+    },
+    {
+      "epoch": 0.16855263157894737,
+      "grad_norm": 3.59375,
+      "grad_norm_var": 0.13404541015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0518,
+      "loss/crossentropy": 2.456227493286133,
+      "loss/hidden": 2.8640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27998181581497195,
+      "loss/reg": 0.0,
+      "step": 25620
+    },
+    {
+      "epoch": 0.16861842105263158,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.15030899047851562,
+      "learning_rate": 0.0001,
+      "loss": 3.043,
+      "loss/crossentropy": 2.208226388692856,
+      "loss/hidden": 2.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24489136189222335,
+      "loss/reg": 0.0,
+      "step": 25630
+    },
+    {
+      "epoch": 0.1686842105263158,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 1.020721181233724,
+      "learning_rate": 0.0001,
+      "loss": 3.1397,
+      "loss/crossentropy": 2.2634316802024843,
+      "loss/hidden": 3.0328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27392966747283937,
+      "loss/reg": 0.0,
+      "step": 25640
+    },
+    {
+      "epoch": 0.16875,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.980279286702474,
+      "learning_rate": 0.0001,
+      "loss": 3.0574,
+      "loss/crossentropy": 2.2914888978004457,
+      "loss/hidden": 2.6984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23568858355283737,
+      "loss/reg": 0.0,
+      "step": 25650
+    },
+    {
+      "epoch": 0.1688157894736842,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.07039769490559895,
+      "learning_rate": 0.0001,
+      "loss": 3.139,
+      "loss/crossentropy": 2.1703743815422056,
+      "loss/hidden": 3.01875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25462436228990554,
+      "loss/reg": 0.0,
+      "step": 25660
+    },
+    {
+      "epoch": 0.1688815789473684,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.22234598795572916,
+      "learning_rate": 0.0001,
+      "loss": 3.0749,
+      "loss/crossentropy": 2.115765154361725,
+      "loss/hidden": 2.9765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2619611322879791,
+      "loss/reg": 0.0,
+      "step": 25670
+    },
+    {
+      "epoch": 0.16894736842105262,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.21928609212239583,
+      "learning_rate": 0.0001,
+      "loss": 3.1309,
+      "loss/crossentropy": 2.355665123462677,
+      "loss/hidden": 2.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25024047791957854,
+      "loss/reg": 0.0,
+      "step": 25680
+    },
+    {
+      "epoch": 0.16901315789473684,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.6276204427083333,
+      "learning_rate": 0.0001,
+      "loss": 3.2325,
+      "loss/crossentropy": 2.3543809175491335,
+      "loss/hidden": 2.90625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28042998909950256,
+      "loss/reg": 0.0,
+      "step": 25690
+    },
+    {
+      "epoch": 0.16907894736842105,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.55308837890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1367,
+      "loss/crossentropy": 2.0251551032066346,
+      "loss/hidden": 2.815625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2671583190560341,
+      "loss/reg": 0.0,
+      "step": 25700
+    },
+    {
+      "epoch": 0.16914473684210526,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.8514638264973958,
+      "learning_rate": 0.0001,
+      "loss": 3.0816,
+      "loss/crossentropy": 2.3947930097579957,
+      "loss/hidden": 2.7453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23190638422966003,
+      "loss/reg": 0.0,
+      "step": 25710
+    },
+    {
+      "epoch": 0.16921052631578948,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.8416341145833334,
+      "learning_rate": 0.0001,
+      "loss": 3.0317,
+      "loss/crossentropy": 2.3030895352363587,
+      "loss/hidden": 2.996875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25654992610216143,
+      "loss/reg": 0.0,
+      "step": 25720
+    },
+    {
+      "epoch": 0.1692763157894737,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.1338043212890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0553,
+      "loss/crossentropy": 2.4580634593963624,
+      "loss/hidden": 2.996875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29121011197566987,
+      "loss/reg": 0.0,
+      "step": 25730
+    },
+    {
+      "epoch": 0.1693421052631579,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.12617162068684895,
+      "learning_rate": 0.0001,
+      "loss": 3.0517,
+      "loss/crossentropy": 2.2136544942855836,
+      "loss/hidden": 2.7203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2460548087954521,
+      "loss/reg": 0.0,
+      "step": 25740
+    },
+    {
+      "epoch": 0.16940789473684212,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.018700154622395833,
+      "learning_rate": 0.0001,
+      "loss": 3.0731,
+      "loss/crossentropy": 2.2090337753295897,
+      "loss/hidden": 2.978125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26342562288045884,
+      "loss/reg": 0.0,
+      "step": 25750
+    },
+    {
+      "epoch": 0.1694736842105263,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.31311442057291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1412,
+      "loss/crossentropy": 2.2381184220314028,
+      "loss/hidden": 3.1265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3542740896344185,
+      "loss/reg": 0.0,
+      "step": 25760
+    },
+    {
+      "epoch": 0.16953947368421052,
+      "grad_norm": 3.9375,
+      "grad_norm_var": 0.39954020182291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1884,
+      "loss/crossentropy": 2.115560531616211,
+      "loss/hidden": 2.9421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2626333341002464,
+      "loss/reg": 0.0,
+      "step": 25770
+    },
+    {
+      "epoch": 0.16960526315789473,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.21519775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0564,
+      "loss/crossentropy": 1.9519969999790192,
+      "loss/hidden": 2.8859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22164231091737746,
+      "loss/reg": 0.0,
+      "step": 25780
+    },
+    {
+      "epoch": 0.16967105263157894,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.08953450520833334,
+      "learning_rate": 0.0001,
+      "loss": 3.1503,
+      "loss/crossentropy": 2.2421632409095764,
+      "loss/hidden": 2.8140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23813997954130173,
+      "loss/reg": 0.0,
+      "step": 25790
+    },
+    {
+      "epoch": 0.16973684210526316,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.1667144775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0487,
+      "loss/crossentropy": 1.9510473489761353,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1950627200305462,
+      "loss/reg": 0.0,
+      "step": 25800
+    },
+    {
+      "epoch": 0.16980263157894737,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.15810139973958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0855,
+      "loss/crossentropy": 2.2411985039710998,
+      "loss/hidden": 2.871875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25535731613636015,
+      "loss/reg": 0.0,
+      "step": 25810
+    },
+    {
+      "epoch": 0.16986842105263159,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 1.0436358133951822,
+      "learning_rate": 0.0001,
+      "loss": 3.115,
+      "loss/crossentropy": 2.045396554470062,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22041565626859666,
+      "loss/reg": 0.0,
+      "step": 25820
+    },
+    {
+      "epoch": 0.1699342105263158,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.18515599568684896,
+      "learning_rate": 0.0001,
+      "loss": 3.0967,
+      "loss/crossentropy": 2.4469074010849,
+      "loss/hidden": 2.871875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2558215782046318,
+      "loss/reg": 0.0,
+      "step": 25830
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.38408203125,
+      "learning_rate": 0.0001,
+      "loss": 3.1042,
+      "loss/crossentropy": 2.038024789094925,
+      "loss/hidden": 2.9796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28011882603168486,
+      "loss/reg": 0.0,
+      "step": 25840
+    },
+    {
+      "epoch": 0.1700657894736842,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.12269694010416667,
+      "learning_rate": 0.0001,
+      "loss": 3.0198,
+      "loss/crossentropy": 2.246475076675415,
+      "loss/hidden": 2.71875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21217122972011565,
+      "loss/reg": 0.0,
+      "step": 25850
+    },
+    {
+      "epoch": 0.1701315789473684,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.5065388997395833,
+      "learning_rate": 0.0001,
+      "loss": 3.0953,
+      "loss/crossentropy": 2.2885293424129487,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23523461371660231,
+      "loss/reg": 0.0,
+      "step": 25860
+    },
+    {
+      "epoch": 0.17019736842105262,
+      "grad_norm": 4.71875,
+      "grad_norm_var": 0.8165323893229167,
+      "learning_rate": 0.0001,
+      "loss": 3.0392,
+      "loss/crossentropy": 2.1574928402900695,
+      "loss/hidden": 2.9,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25582125633955,
+      "loss/reg": 0.0,
+      "step": 25870
+    },
+    {
+      "epoch": 0.17026315789473684,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.57008056640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1252,
+      "loss/crossentropy": 2.563046908378601,
+      "loss/hidden": 3.2078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2562621384859085,
+      "loss/reg": 0.0,
+      "step": 25880
+    },
+    {
+      "epoch": 0.17032894736842105,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.2592437744140625,
+      "learning_rate": 0.0001,
+      "loss": 3.054,
+      "loss/crossentropy": 2.4634771227836607,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25339619666337965,
+      "loss/reg": 0.0,
+      "step": 25890
+    },
+    {
+      "epoch": 0.17039473684210527,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.04925130208333333,
+      "learning_rate": 0.0001,
+      "loss": 3.0529,
+      "loss/crossentropy": 2.5104726552963257,
+      "loss/hidden": 2.8765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24876073151826858,
+      "loss/reg": 0.0,
+      "step": 25900
+    },
+    {
+      "epoch": 0.17046052631578948,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.07171122233072917,
+      "learning_rate": 0.0001,
+      "loss": 3.0629,
+      "loss/crossentropy": 2.2496933460235597,
+      "loss/hidden": 2.734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2183684565126896,
+      "loss/reg": 0.0,
+      "step": 25910
+    },
+    {
+      "epoch": 0.1705263157894737,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.05900777180989583,
+      "learning_rate": 0.0001,
+      "loss": 3.0543,
+      "loss/crossentropy": 2.3338397264480593,
+      "loss/hidden": 2.9109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26187729388475417,
+      "loss/reg": 0.0,
+      "step": 25920
+    },
+    {
+      "epoch": 0.1705921052631579,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.018359375,
+      "learning_rate": 0.0001,
+      "loss": 3.0801,
+      "loss/crossentropy": 2.499437928199768,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25035695284605025,
+      "loss/reg": 0.0,
+      "step": 25930
+    },
+    {
+      "epoch": 0.1706578947368421,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.014264933268229167,
+      "learning_rate": 0.0001,
+      "loss": 3.0337,
+      "loss/crossentropy": 2.217645859718323,
+      "loss/hidden": 2.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21889251619577407,
+      "loss/reg": 0.0,
+      "step": 25940
+    },
+    {
+      "epoch": 0.1707236842105263,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.04052734375,
+      "learning_rate": 0.0001,
+      "loss": 2.9938,
+      "loss/crossentropy": 2.371463644504547,
+      "loss/hidden": 2.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2288275495171547,
+      "loss/reg": 0.0,
+      "step": 25950
+    },
+    {
+      "epoch": 0.17078947368421052,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.1755859375,
+      "learning_rate": 0.0001,
+      "loss": 3.0916,
+      "loss/crossentropy": 2.0315194338560105,
+      "loss/hidden": 3.028125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23134928867220877,
+      "loss/reg": 0.0,
+      "step": 25960
+    },
+    {
+      "epoch": 0.17085526315789473,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.18121744791666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0837,
+      "loss/crossentropy": 2.1368157267570496,
+      "loss/hidden": 2.8765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2233546957373619,
+      "loss/reg": 0.0,
+      "step": 25970
+    },
+    {
+      "epoch": 0.17092105263157895,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.027311197916666665,
+      "learning_rate": 0.0001,
+      "loss": 3.0984,
+      "loss/crossentropy": 2.3421939969062806,
+      "loss/hidden": 3.040625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28762088268995284,
+      "loss/reg": 0.0,
+      "step": 25980
+    },
+    {
+      "epoch": 0.17098684210526316,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.051656087239583336,
+      "learning_rate": 0.0001,
+      "loss": 3.0788,
+      "loss/crossentropy": 2.002690005302429,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24228747338056564,
+      "loss/reg": 0.0,
+      "step": 25990
+    },
+    {
+      "epoch": 0.17105263157894737,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.049397786458333336,
+      "learning_rate": 0.0001,
+      "loss": 3.0917,
+      "loss/crossentropy": 2.6486097812652587,
+      "loss/hidden": 2.7515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24952505975961686,
+      "loss/reg": 0.0,
+      "step": 26000
+    },
+    {
+      "epoch": 0.1711184210526316,
+      "grad_norm": 2.0,
+      "grad_norm_var": 0.0459869384765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0288,
+      "loss/crossentropy": 2.516502094268799,
+      "loss/hidden": 2.703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24382732957601547,
+      "loss/reg": 0.0,
+      "step": 26010
+    },
+    {
+      "epoch": 0.1711842105263158,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.04586181640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0415,
+      "loss/crossentropy": 2.171498954296112,
+      "loss/hidden": 2.65625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20375996455550194,
+      "loss/reg": 0.0,
+      "step": 26020
+    },
+    {
+      "epoch": 0.17125,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.0759600321451823,
+      "learning_rate": 0.0001,
+      "loss": 3.0905,
+      "loss/crossentropy": 2.2978348255157472,
+      "loss/hidden": 2.94375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2875170633196831,
+      "loss/reg": 0.0,
+      "step": 26030
+    },
+    {
+      "epoch": 0.1713157894736842,
+      "grad_norm": 3.546875,
+      "grad_norm_var": 0.1603167215983073,
+      "learning_rate": 0.0001,
+      "loss": 3.0384,
+      "loss/crossentropy": 2.462969756126404,
+      "loss/hidden": 2.8390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2628189787268639,
+      "loss/reg": 0.0,
+      "step": 26040
+    },
+    {
+      "epoch": 0.1713815789473684,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.13243815104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0513,
+      "loss/crossentropy": 2.537035143375397,
+      "loss/hidden": 2.765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24593261033296585,
+      "loss/reg": 0.0,
+      "step": 26050
+    },
+    {
+      "epoch": 0.17144736842105263,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.061102040608723956,
+      "learning_rate": 0.0001,
+      "loss": 3.0445,
+      "loss/crossentropy": 2.2031071186065674,
+      "loss/hidden": 2.884375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24524077475070954,
+      "loss/reg": 0.0,
+      "step": 26060
+    },
+    {
+      "epoch": 0.17151315789473684,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.009764607747395833,
+      "learning_rate": 0.0001,
+      "loss": 3.0725,
+      "loss/crossentropy": 2.4979568481445313,
+      "loss/hidden": 2.709375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2647134616971016,
+      "loss/reg": 0.0,
+      "step": 26070
+    },
+    {
+      "epoch": 0.17157894736842105,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.026024373372395833,
+      "learning_rate": 0.0001,
+      "loss": 2.99,
+      "loss/crossentropy": 2.350401425361633,
+      "loss/hidden": 2.78125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24495915323495865,
+      "loss/reg": 0.0,
+      "step": 26080
+    },
+    {
+      "epoch": 0.17164473684210527,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.028857421875,
+      "learning_rate": 0.0001,
+      "loss": 3.0236,
+      "loss/crossentropy": 2.275194299221039,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22880630493164061,
+      "loss/reg": 0.0,
+      "step": 26090
+    },
+    {
+      "epoch": 0.17171052631578948,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.03967692057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0823,
+      "loss/crossentropy": 2.000411808490753,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2721859060227871,
+      "loss/reg": 0.0,
+      "step": 26100
+    },
+    {
+      "epoch": 0.1717763157894737,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.06353759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0778,
+      "loss/crossentropy": 2.305191624164581,
+      "loss/hidden": 2.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2284097969532013,
+      "loss/reg": 0.0,
+      "step": 26110
+    },
+    {
+      "epoch": 0.1718421052631579,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.34947509765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0215,
+      "loss/crossentropy": 2.3149521112442017,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27900824695825577,
+      "loss/reg": 0.0,
+      "step": 26120
+    },
+    {
+      "epoch": 0.1719078947368421,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.39952977498372394,
+      "learning_rate": 0.0001,
+      "loss": 3.0197,
+      "loss/crossentropy": 2.421990168094635,
+      "loss/hidden": 2.8265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2581001713871956,
+      "loss/reg": 0.0,
+      "step": 26130
+    },
+    {
+      "epoch": 0.1719736842105263,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.19842529296875,
+      "learning_rate": 0.0001,
+      "loss": 3.115,
+      "loss/crossentropy": 2.473453497886658,
+      "loss/hidden": 2.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25612604022026064,
+      "loss/reg": 0.0,
+      "step": 26140
+    },
+    {
+      "epoch": 0.17203947368421052,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.06412328084309896,
+      "learning_rate": 0.0001,
+      "loss": 3.1205,
+      "loss/crossentropy": 2.395763027667999,
+      "loss/hidden": 2.8953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25399915128946304,
+      "loss/reg": 0.0,
+      "step": 26150
+    },
+    {
+      "epoch": 0.17210526315789473,
+      "grad_norm": 3.21875,
+      "grad_norm_var": 3.7263832092285156,
+      "learning_rate": 0.0001,
+      "loss": 3.1443,
+      "loss/crossentropy": 2.3229804635047913,
+      "loss/hidden": 2.890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25968860387802123,
+      "loss/reg": 0.0,
+      "step": 26160
+    },
+    {
+      "epoch": 0.17217105263157895,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 3.5683990478515626,
+      "learning_rate": 0.0001,
+      "loss": 3.0833,
+      "loss/crossentropy": 2.27206437587738,
+      "loss/hidden": 2.9453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26493641585111616,
+      "loss/reg": 0.0,
+      "step": 26170
+    },
+    {
+      "epoch": 0.17223684210526316,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.140771484375,
+      "learning_rate": 0.0001,
+      "loss": 3.1236,
+      "loss/crossentropy": 2.1949939370155334,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24421997666358947,
+      "loss/reg": 0.0,
+      "step": 26180
+    },
+    {
+      "epoch": 0.17230263157894737,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.08112691243489584,
+      "learning_rate": 0.0001,
+      "loss": 3.0485,
+      "loss/crossentropy": 2.292436492443085,
+      "loss/hidden": 2.66875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2527725785970688,
+      "loss/reg": 0.0,
+      "step": 26190
+    },
+    {
+      "epoch": 0.1723684210526316,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.08532613118489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1316,
+      "loss/crossentropy": 2.497014009952545,
+      "loss/hidden": 2.8265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2447853922843933,
+      "loss/reg": 0.0,
+      "step": 26200
+    },
+    {
+      "epoch": 0.1724342105263158,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.142578125,
+      "learning_rate": 0.0001,
+      "loss": 3.1014,
+      "loss/crossentropy": 2.22248170375824,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2591711387038231,
+      "loss/reg": 0.0,
+      "step": 26210
+    },
+    {
+      "epoch": 0.1725,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.12844009399414064,
+      "learning_rate": 0.0001,
+      "loss": 3.0061,
+      "loss/crossentropy": 2.3277372121810913,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2391911566257477,
+      "loss/reg": 0.0,
+      "step": 26220
+    },
+    {
+      "epoch": 0.1725657894736842,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.03429361979166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0468,
+      "loss/crossentropy": 2.593550610542297,
+      "loss/hidden": 2.69375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26268754005432127,
+      "loss/reg": 0.0,
+      "step": 26230
+    },
+    {
+      "epoch": 0.1726315789473684,
+      "grad_norm": 3.59375,
+      "grad_norm_var": 0.13815689086914062,
+      "learning_rate": 0.0001,
+      "loss": 3.0785,
+      "loss/crossentropy": 2.037731957435608,
+      "loss/hidden": 2.7046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19823871441185476,
+      "loss/reg": 0.0,
+      "step": 26240
+    },
+    {
+      "epoch": 0.17269736842105263,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.1658404032389323,
+      "learning_rate": 0.0001,
+      "loss": 3.0753,
+      "loss/crossentropy": 2.365402173995972,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22439467608928682,
+      "loss/reg": 0.0,
+      "step": 26250
+    },
+    {
+      "epoch": 0.17276315789473684,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.42582906087239586,
+      "learning_rate": 0.0001,
+      "loss": 3.0247,
+      "loss/crossentropy": 2.4660609483718874,
+      "loss/hidden": 2.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2206026643514633,
+      "loss/reg": 0.0,
+      "step": 26260
+    },
+    {
+      "epoch": 0.17282894736842105,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.4584307352701823,
+      "learning_rate": 0.0001,
+      "loss": 2.9979,
+      "loss/crossentropy": 2.3537610173225403,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24690964370965957,
+      "loss/reg": 0.0,
+      "step": 26270
+    },
+    {
+      "epoch": 0.17289473684210527,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.1235015869140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0554,
+      "loss/crossentropy": 2.4086636185646055,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22505185902118682,
+      "loss/reg": 0.0,
+      "step": 26280
+    },
+    {
+      "epoch": 0.17296052631578948,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.26606343587239584,
+      "learning_rate": 0.0001,
+      "loss": 3.0265,
+      "loss/crossentropy": 2.120291304588318,
+      "loss/hidden": 2.725,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20824237614870073,
+      "loss/reg": 0.0,
+      "step": 26290
+    },
+    {
+      "epoch": 0.1730263157894737,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.18574117024739584,
+      "learning_rate": 0.0001,
+      "loss": 2.9919,
+      "loss/crossentropy": 2.1023782581090926,
+      "loss/hidden": 2.6734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19826814979314805,
+      "loss/reg": 0.0,
+      "step": 26300
+    },
+    {
+      "epoch": 0.1730921052631579,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.24006754557291668,
+      "learning_rate": 0.0001,
+      "loss": 3.1028,
+      "loss/crossentropy": 2.226356017589569,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2514319851994514,
+      "loss/reg": 0.0,
+      "step": 26310
+    },
+    {
+      "epoch": 0.1731578947368421,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.234326171875,
+      "learning_rate": 0.0001,
+      "loss": 3.0524,
+      "loss/crossentropy": 2.2389047503471375,
+      "loss/hidden": 2.725,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23142587393522263,
+      "loss/reg": 0.0,
+      "step": 26320
+    },
+    {
+      "epoch": 0.1732236842105263,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.07327041625976563,
+      "learning_rate": 0.0001,
+      "loss": 3.054,
+      "loss/crossentropy": 2.3862741947174073,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2310537427663803,
+      "loss/reg": 0.0,
+      "step": 26330
+    },
+    {
+      "epoch": 0.17328947368421052,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.07641499837239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0653,
+      "loss/crossentropy": 2.348588991165161,
+      "loss/hidden": 2.678125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22305506616830825,
+      "loss/reg": 0.0,
+      "step": 26340
+    },
+    {
+      "epoch": 0.17335526315789473,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.05279541015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0529,
+      "loss/crossentropy": 2.036811423301697,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23410281240940095,
+      "loss/reg": 0.0,
+      "step": 26350
+    },
+    {
+      "epoch": 0.17342105263157895,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.2839101155598958,
+      "learning_rate": 0.0001,
+      "loss": 3.1092,
+      "loss/crossentropy": 2.1091545104980467,
+      "loss/hidden": 2.9046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23594892621040345,
+      "loss/reg": 0.0,
+      "step": 26360
+    },
+    {
+      "epoch": 0.17348684210526316,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.061848958333333336,
+      "learning_rate": 0.0001,
+      "loss": 3.091,
+      "loss/crossentropy": 2.290253794193268,
+      "loss/hidden": 2.8625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29689010232686996,
+      "loss/reg": 0.0,
+      "step": 26370
+    },
+    {
+      "epoch": 0.17355263157894738,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.04799702962239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0457,
+      "loss/crossentropy": 2.193889284133911,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22776648998260499,
+      "loss/reg": 0.0,
+      "step": 26380
+    },
+    {
+      "epoch": 0.1736184210526316,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.08166910807291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0939,
+      "loss/crossentropy": 2.2826230049133303,
+      "loss/hidden": 2.9390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2337636888027191,
+      "loss/reg": 0.0,
+      "step": 26390
+    },
+    {
+      "epoch": 0.1736842105263158,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.16920572916666668,
+      "learning_rate": 0.0001,
+      "loss": 3.0459,
+      "loss/crossentropy": 2.233789896965027,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20103665292263032,
+      "loss/reg": 0.0,
+      "step": 26400
+    },
+    {
+      "epoch": 0.17375,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.17185440063476562,
+      "learning_rate": 0.0001,
+      "loss": 3.0538,
+      "loss/crossentropy": 2.295468533039093,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2207227498292923,
+      "loss/reg": 0.0,
+      "step": 26410
+    },
+    {
+      "epoch": 0.1738157894736842,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.048618316650390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0468,
+      "loss/crossentropy": 2.405815064907074,
+      "loss/hidden": 2.7046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22815033197402954,
+      "loss/reg": 0.0,
+      "step": 26420
+    },
+    {
+      "epoch": 0.17388157894736841,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.05583394368489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1146,
+      "loss/crossentropy": 2.2692611694335936,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24037961512804032,
+      "loss/reg": 0.0,
+      "step": 26430
+    },
+    {
+      "epoch": 0.17394736842105263,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.1251617431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1101,
+      "loss/crossentropy": 2.2592686772346497,
+      "loss/hidden": 3.0109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24446354508399964,
+      "loss/reg": 0.0,
+      "step": 26440
+    },
+    {
+      "epoch": 0.17401315789473684,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.10937093098958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0527,
+      "loss/crossentropy": 2.291548252105713,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25084199756383896,
+      "loss/reg": 0.0,
+      "step": 26450
+    },
+    {
+      "epoch": 0.17407894736842106,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 0.06529947916666666,
+      "learning_rate": 0.0001,
+      "loss": 3.1052,
+      "loss/crossentropy": 2.37350834608078,
+      "loss/hidden": 2.8890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2556142807006836,
+      "loss/reg": 0.0,
+      "step": 26460
+    },
+    {
+      "epoch": 0.17414473684210527,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.07598037719726562,
+      "learning_rate": 0.0001,
+      "loss": 3.0113,
+      "loss/crossentropy": 2.250531017780304,
+      "loss/hidden": 2.8453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2357124164700508,
+      "loss/reg": 0.0,
+      "step": 26470
+    },
+    {
+      "epoch": 0.17421052631578948,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.17610677083333334,
+      "learning_rate": 0.0001,
+      "loss": 3.1412,
+      "loss/crossentropy": 2.439283573627472,
+      "loss/hidden": 2.6515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2121993750333786,
+      "loss/reg": 0.0,
+      "step": 26480
+    },
+    {
+      "epoch": 0.1742763157894737,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.16172587076822917,
+      "learning_rate": 0.0001,
+      "loss": 3.0562,
+      "loss/crossentropy": 2.3756911277771,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2490501657128334,
+      "loss/reg": 0.0,
+      "step": 26490
+    },
+    {
+      "epoch": 0.17434210526315788,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.050211588541666664,
+      "learning_rate": 0.0001,
+      "loss": 3.0538,
+      "loss/crossentropy": 2.5375622749328612,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24366173297166824,
+      "loss/reg": 0.0,
+      "step": 26500
+    },
+    {
+      "epoch": 0.1744078947368421,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.5032185872395833,
+      "learning_rate": 0.0001,
+      "loss": 3.0919,
+      "loss/crossentropy": 2.3419446110725404,
+      "loss/hidden": 2.9359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2749262437224388,
+      "loss/reg": 0.0,
+      "step": 26510
+    },
+    {
+      "epoch": 0.1744736842105263,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.03943583170572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0832,
+      "loss/crossentropy": 2.070459759235382,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2193553477525711,
+      "loss/reg": 0.0,
+      "step": 26520
+    },
+    {
+      "epoch": 0.17453947368421052,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.11748758951822917,
+      "learning_rate": 0.0001,
+      "loss": 3.0301,
+      "loss/crossentropy": 2.2101051688194273,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2358426868915558,
+      "loss/reg": 0.0,
+      "step": 26530
+    },
+    {
+      "epoch": 0.17460526315789474,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.11904195149739584,
+      "learning_rate": 0.0001,
+      "loss": 2.9898,
+      "loss/crossentropy": 2.301384377479553,
+      "loss/hidden": 2.9671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.270483261346817,
+      "loss/reg": 0.0,
+      "step": 26540
+    },
+    {
+      "epoch": 0.17467105263157895,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.03364969889322917,
+      "learning_rate": 0.0001,
+      "loss": 2.9894,
+      "loss/crossentropy": 2.233541655540466,
+      "loss/hidden": 2.7265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23210279494524003,
+      "loss/reg": 0.0,
+      "step": 26550
+    },
+    {
+      "epoch": 0.17473684210526316,
+      "grad_norm": 4.3125,
+      "grad_norm_var": 0.30920817057291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0375,
+      "loss/crossentropy": 2.4949169993400573,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27407604902982713,
+      "loss/reg": 0.0,
+      "step": 26560
+    },
+    {
+      "epoch": 0.17480263157894738,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.4443511962890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1064,
+      "loss/crossentropy": 2.180380403995514,
+      "loss/hidden": 3.06875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2965380325913429,
+      "loss/reg": 0.0,
+      "step": 26570
+    },
+    {
+      "epoch": 0.1748684210526316,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.12886962890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0636,
+      "loss/crossentropy": 2.2048864006996154,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24606209397315978,
+      "loss/reg": 0.0,
+      "step": 26580
+    },
+    {
+      "epoch": 0.1749342105263158,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.090576171875,
+      "learning_rate": 0.0001,
+      "loss": 3.0745,
+      "loss/crossentropy": 2.0117301523685454,
+      "loss/hidden": 2.815625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.211642824113369,
+      "loss/reg": 0.0,
+      "step": 26590
+    },
+    {
+      "epoch": 0.175,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.9207834879557292,
+      "learning_rate": 0.0001,
+      "loss": 2.9913,
+      "loss/crossentropy": 2.2407095432281494,
+      "loss/hidden": 2.8203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25644198805093765,
+      "loss/reg": 0.0,
+      "step": 26600
+    },
+    {
+      "epoch": 0.1750657894736842,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.9292063395182292,
+      "learning_rate": 0.0001,
+      "loss": 3.1122,
+      "loss/crossentropy": 2.510072946548462,
+      "loss/hidden": 2.821875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24346872568130493,
+      "loss/reg": 0.0,
+      "step": 26610
+    },
+    {
+      "epoch": 0.17513157894736842,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 1.2361806233723958,
+      "learning_rate": 0.0001,
+      "loss": 3.1052,
+      "loss/crossentropy": 2.0434396266937256,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19786725342273712,
+      "loss/reg": 0.0,
+      "step": 26620
+    },
+    {
+      "epoch": 0.17519736842105263,
+      "grad_norm": 2.25,
+      "grad_norm_var": 1.2163075764973958,
+      "learning_rate": 0.0001,
+      "loss": 3.1224,
+      "loss/crossentropy": 2.2104805946350097,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21630636900663375,
+      "loss/reg": 0.0,
+      "step": 26630
+    },
+    {
+      "epoch": 0.17526315789473684,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.03358739217122396,
+      "learning_rate": 0.0001,
+      "loss": 3.028,
+      "loss/crossentropy": 2.178933525085449,
+      "loss/hidden": 2.6953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20190188586711882,
+      "loss/reg": 0.0,
+      "step": 26640
+    },
+    {
+      "epoch": 0.17532894736842106,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.09688212076822916,
+      "learning_rate": 0.0001,
+      "loss": 3.0897,
+      "loss/crossentropy": 2.2301127552986144,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23655431792140008,
+      "loss/reg": 0.0,
+      "step": 26650
+    },
+    {
+      "epoch": 0.17539473684210527,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.10475972493489584,
+      "learning_rate": 0.0001,
+      "loss": 3.0935,
+      "loss/crossentropy": 2.2094313383102415,
+      "loss/hidden": 2.9734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27910117208957674,
+      "loss/reg": 0.0,
+      "step": 26660
+    },
+    {
+      "epoch": 0.17546052631578948,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.050446573893229166,
+      "learning_rate": 0.0001,
+      "loss": 3.0452,
+      "loss/crossentropy": 2.491715204715729,
+      "loss/hidden": 2.7171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24501692056655883,
+      "loss/reg": 0.0,
+      "step": 26670
+    },
+    {
+      "epoch": 0.1755263157894737,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.42122294108072916,
+      "learning_rate": 0.0001,
+      "loss": 3.1247,
+      "loss/crossentropy": 2.431166636943817,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24856783598661422,
+      "loss/reg": 0.0,
+      "step": 26680
+    },
+    {
+      "epoch": 0.17559210526315788,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.0764801025390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1265,
+      "loss/crossentropy": 2.2466578841209413,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22143050357699395,
+      "loss/reg": 0.0,
+      "step": 26690
+    },
+    {
+      "epoch": 0.1756578947368421,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.06314697265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0522,
+      "loss/crossentropy": 1.8643316149711608,
+      "loss/hidden": 2.8640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22551749348640443,
+      "loss/reg": 0.0,
+      "step": 26700
+    },
+    {
+      "epoch": 0.1757236842105263,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.12968648274739583,
+      "learning_rate": 0.0001,
+      "loss": 3.0386,
+      "loss/crossentropy": 1.784457266330719,
+      "loss/hidden": 2.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.17761861719191074,
+      "loss/reg": 0.0,
+      "step": 26710
+    },
+    {
+      "epoch": 0.17578947368421052,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.13782450358072917,
+      "learning_rate": 0.0001,
+      "loss": 3.0727,
+      "loss/crossentropy": 2.388060462474823,
+      "loss/hidden": 2.9546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25690646171569825,
+      "loss/reg": 0.0,
+      "step": 26720
+    },
+    {
+      "epoch": 0.17585526315789474,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.5772450764973959,
+      "learning_rate": 0.0001,
+      "loss": 3.017,
+      "loss/crossentropy": 2.293041491508484,
+      "loss/hidden": 2.9359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2976214215159416,
+      "loss/reg": 0.0,
+      "step": 26730
+    },
+    {
+      "epoch": 0.17592105263157895,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.7439737955729167,
+      "learning_rate": 0.0001,
+      "loss": 3.1688,
+      "loss/crossentropy": 2.346044087409973,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24346352070569993,
+      "loss/reg": 0.0,
+      "step": 26740
+    },
+    {
+      "epoch": 0.17598684210526316,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.19332275390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1074,
+      "loss/crossentropy": 2.3460346817970277,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22802013754844666,
+      "loss/reg": 0.0,
+      "step": 26750
+    },
+    {
+      "epoch": 0.17605263157894738,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.09132486979166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0194,
+      "loss/crossentropy": 2.361405539512634,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2546524882316589,
+      "loss/reg": 0.0,
+      "step": 26760
+    },
+    {
+      "epoch": 0.1761184210526316,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.06636962890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1365,
+      "loss/crossentropy": 2.184977853298187,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24620451256632805,
+      "loss/reg": 0.0,
+      "step": 26770
+    },
+    {
+      "epoch": 0.17618421052631578,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.04890034993489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0491,
+      "loss/crossentropy": 2.3393765091896057,
+      "loss/hidden": 2.7234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21822300404310227,
+      "loss/reg": 0.0,
+      "step": 26780
+    },
+    {
+      "epoch": 0.17625,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.08136571248372396,
+      "learning_rate": 0.0001,
+      "loss": 3.0288,
+      "loss/crossentropy": 2.1801982522010803,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2308456853032112,
+      "loss/reg": 0.0,
+      "step": 26790
+    },
+    {
+      "epoch": 0.1763157894736842,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.08675918579101563,
+      "learning_rate": 0.0001,
+      "loss": 3.0063,
+      "loss/crossentropy": 2.3970799446105957,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23122948557138442,
+      "loss/reg": 0.0,
+      "step": 26800
+    },
+    {
+      "epoch": 0.17638157894736842,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.04929911295572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0628,
+      "loss/crossentropy": 2.3741040945053102,
+      "loss/hidden": 2.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26243693232536314,
+      "loss/reg": 0.0,
+      "step": 26810
+    },
+    {
+      "epoch": 0.17644736842105263,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.07210184733072916,
+      "learning_rate": 0.0001,
+      "loss": 3.0187,
+      "loss/crossentropy": 2.2319773197174073,
+      "loss/hidden": 2.66875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.201442664116621,
+      "loss/reg": 0.0,
+      "step": 26820
+    },
+    {
+      "epoch": 0.17651315789473684,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.20078023274739584,
+      "learning_rate": 0.0001,
+      "loss": 3.0845,
+      "loss/crossentropy": 2.0902840554714204,
+      "loss/hidden": 2.91875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23434632122516633,
+      "loss/reg": 0.0,
+      "step": 26830
+    },
+    {
+      "epoch": 0.17657894736842106,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.48001200358072915,
+      "learning_rate": 0.0001,
+      "loss": 3.0548,
+      "loss/crossentropy": 1.8517399907112122,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1863703802227974,
+      "loss/reg": 0.0,
+      "step": 26840
+    },
+    {
+      "epoch": 0.17664473684210527,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.4188313802083333,
+      "learning_rate": 0.0001,
+      "loss": 3.0383,
+      "loss/crossentropy": 2.398709797859192,
+      "loss/hidden": 2.65625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2097424551844597,
+      "loss/reg": 0.0,
+      "step": 26850
+    },
+    {
+      "epoch": 0.17671052631578948,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.14855855305989582,
+      "learning_rate": 0.0001,
+      "loss": 3.0439,
+      "loss/crossentropy": 2.2976414561271667,
+      "loss/hidden": 2.659375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2062876433134079,
+      "loss/reg": 0.0,
+      "step": 26860
+    },
+    {
+      "epoch": 0.1767763157894737,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.3957801818847656,
+      "learning_rate": 0.0001,
+      "loss": 3.1221,
+      "loss/crossentropy": 2.2302059173583983,
+      "loss/hidden": 3.04375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2583988204598427,
+      "loss/reg": 0.0,
+      "step": 26870
+    },
+    {
+      "epoch": 0.17684210526315788,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.34606831868489585,
+      "learning_rate": 0.0001,
+      "loss": 3.0334,
+      "loss/crossentropy": 2.2456519961357118,
+      "loss/hidden": 2.890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26051086038351057,
+      "loss/reg": 0.0,
+      "step": 26880
+    },
+    {
+      "epoch": 0.1769078947368421,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.09192301432291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0806,
+      "loss/crossentropy": 2.25178804397583,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2424270063638687,
+      "loss/reg": 0.0,
+      "step": 26890
+    },
+    {
+      "epoch": 0.1769736842105263,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.179736328125,
+      "learning_rate": 0.0001,
+      "loss": 3.0299,
+      "loss/crossentropy": 2.2696429252624513,
+      "loss/hidden": 2.95,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25682739466428756,
+      "loss/reg": 0.0,
+      "step": 26900
+    },
+    {
+      "epoch": 0.17703947368421052,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.44387613932291664,
+      "learning_rate": 0.0001,
+      "loss": 3.0266,
+      "loss/crossentropy": 2.349669575691223,
+      "loss/hidden": 2.84375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24149384647607802,
+      "loss/reg": 0.0,
+      "step": 26910
+    },
+    {
+      "epoch": 0.17710526315789474,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.34541727701822916,
+      "learning_rate": 0.0001,
+      "loss": 3.0779,
+      "loss/crossentropy": 2.20119423866272,
+      "loss/hidden": 2.6890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21436720862984657,
+      "loss/reg": 0.0,
+      "step": 26920
+    },
+    {
+      "epoch": 0.17717105263157895,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.05929361979166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0616,
+      "loss/crossentropy": 2.2472872853279116,
+      "loss/hidden": 2.7375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22420328930020333,
+      "loss/reg": 0.0,
+      "step": 26930
+    },
+    {
+      "epoch": 0.17723684210526316,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.08504231770833333,
+      "learning_rate": 0.0001,
+      "loss": 3.1167,
+      "loss/crossentropy": 2.308558535575867,
+      "loss/hidden": 3.084375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2697313494980335,
+      "loss/reg": 0.0,
+      "step": 26940
+    },
+    {
+      "epoch": 0.17730263157894738,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.0501617431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0375,
+      "loss/crossentropy": 2.4767362475395203,
+      "loss/hidden": 2.7453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22490220367908478,
+      "loss/reg": 0.0,
+      "step": 26950
+    },
+    {
+      "epoch": 0.1773684210526316,
+      "grad_norm": 4.25,
+      "grad_norm_var": 0.29531962076822915,
+      "learning_rate": 0.0001,
+      "loss": 3.0694,
+      "loss/crossentropy": 2.041945827007294,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21851452216506004,
+      "loss/reg": 0.0,
+      "step": 26960
+    },
+    {
+      "epoch": 0.17743421052631578,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.2590728759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0748,
+      "loss/crossentropy": 2.3148224115371705,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24593613892793656,
+      "loss/reg": 0.0,
+      "step": 26970
+    },
+    {
+      "epoch": 0.1775,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.09059956868489584,
+      "learning_rate": 0.0001,
+      "loss": 3.0916,
+      "loss/crossentropy": 2.3537511348724367,
+      "loss/hidden": 2.7375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26722831577062606,
+      "loss/reg": 0.0,
+      "step": 26980
+    },
+    {
+      "epoch": 0.1775657894736842,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.09703776041666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0801,
+      "loss/crossentropy": 2.3036428809165956,
+      "loss/hidden": 3.0046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28045037388801575,
+      "loss/reg": 0.0,
+      "step": 26990
+    },
+    {
+      "epoch": 0.17763157894736842,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.0448883056640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1064,
+      "loss/crossentropy": 2.210177004337311,
+      "loss/hidden": 2.653125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20902514159679414,
+      "loss/reg": 0.0,
+      "step": 27000
+    },
+    {
+      "epoch": 0.17769736842105263,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.408251953125,
+      "learning_rate": 0.0001,
+      "loss": 3.1257,
+      "loss/crossentropy": 2.354000687599182,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2463086098432541,
+      "loss/reg": 0.0,
+      "step": 27010
+    },
+    {
+      "epoch": 0.17776315789473685,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.07931086222330729,
+      "learning_rate": 0.0001,
+      "loss": 2.9938,
+      "loss/crossentropy": 2.2460681438446044,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2801851168274879,
+      "loss/reg": 0.0,
+      "step": 27020
+    },
+    {
+      "epoch": 0.17782894736842106,
+      "grad_norm": 1.984375,
+      "grad_norm_var": 0.07854715983072917,
+      "learning_rate": 0.0001,
+      "loss": 3.0282,
+      "loss/crossentropy": 2.1731944918632506,
+      "loss/hidden": 2.940625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22212435267865657,
+      "loss/reg": 0.0,
+      "step": 27030
+    },
+    {
+      "epoch": 0.17789473684210527,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.116162109375,
+      "learning_rate": 0.0001,
+      "loss": 3.0344,
+      "loss/crossentropy": 2.441677284240723,
+      "loss/hidden": 2.85625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25195191353559493,
+      "loss/reg": 0.0,
+      "step": 27040
+    },
+    {
+      "epoch": 0.17796052631578949,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.15789388020833334,
+      "learning_rate": 0.0001,
+      "loss": 3.0824,
+      "loss/crossentropy": 2.30471476316452,
+      "loss/hidden": 3.0046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2780482307076454,
+      "loss/reg": 0.0,
+      "step": 27050
+    },
+    {
+      "epoch": 0.17802631578947367,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.13613179524739583,
+      "learning_rate": 0.0001,
+      "loss": 3.0864,
+      "loss/crossentropy": 2.1705368638038633,
+      "loss/hidden": 2.93125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2402413785457611,
+      "loss/reg": 0.0,
+      "step": 27060
+    },
+    {
+      "epoch": 0.17809210526315788,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.09636128743489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1186,
+      "loss/crossentropy": 2.1933916807174683,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22591635286808015,
+      "loss/reg": 0.0,
+      "step": 27070
+    },
+    {
+      "epoch": 0.1781578947368421,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.08137613932291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0876,
+      "loss/crossentropy": 2.2060605615377424,
+      "loss/hidden": 2.9953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2600886657834053,
+      "loss/reg": 0.0,
+      "step": 27080
+    },
+    {
+      "epoch": 0.1782236842105263,
+      "grad_norm": 3.0,
+      "grad_norm_var": 0.12360738118489584,
+      "learning_rate": 0.0001,
+      "loss": 3.0018,
+      "loss/crossentropy": 2.3198684811592103,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23098773807287215,
+      "loss/reg": 0.0,
+      "step": 27090
+    },
+    {
+      "epoch": 0.17828947368421053,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.10364583333333334,
+      "learning_rate": 0.0001,
+      "loss": 2.977,
+      "loss/crossentropy": 2.3410441994667055,
+      "loss/hidden": 2.93125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25723920315504073,
+      "loss/reg": 0.0,
+      "step": 27100
+    },
+    {
+      "epoch": 0.17835526315789474,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.08136393229166666,
+      "learning_rate": 0.0001,
+      "loss": 2.9812,
+      "loss/crossentropy": 2.290076696872711,
+      "loss/hidden": 2.6875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20929899364709853,
+      "loss/reg": 0.0,
+      "step": 27110
+    },
+    {
+      "epoch": 0.17842105263157895,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.061310831705729166,
+      "learning_rate": 0.0001,
+      "loss": 3.0998,
+      "loss/crossentropy": 2.6890960454940798,
+      "loss/hidden": 2.7390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25387548208236693,
+      "loss/reg": 0.0,
+      "step": 27120
+    },
+    {
+      "epoch": 0.17848684210526317,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.03313395182291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0312,
+      "loss/crossentropy": 2.4378479957580566,
+      "loss/hidden": 2.725,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22957875877618789,
+      "loss/reg": 0.0,
+      "step": 27130
+    },
+    {
+      "epoch": 0.17855263157894738,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.11809794108072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1739,
+      "loss/crossentropy": 2.2034701079130175,
+      "loss/hidden": 2.8671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2694556161761284,
+      "loss/reg": 0.0,
+      "step": 27140
+    },
+    {
+      "epoch": 0.17861842105263157,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.243505859375,
+      "learning_rate": 0.0001,
+      "loss": 3.0594,
+      "loss/crossentropy": 2.371540868282318,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23407552391290665,
+      "loss/reg": 0.0,
+      "step": 27150
+    },
+    {
+      "epoch": 0.17868421052631578,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.15914306640625,
+      "learning_rate": 0.0001,
+      "loss": 2.9942,
+      "loss/crossentropy": 1.9778084814548493,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19847352504730226,
+      "loss/reg": 0.0,
+      "step": 27160
+    },
+    {
+      "epoch": 0.17875,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.05803120930989583,
+      "learning_rate": 0.0001,
+      "loss": 3.0767,
+      "loss/crossentropy": 2.3956029176712037,
+      "loss/hidden": 2.69375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21606186181306838,
+      "loss/reg": 0.0,
+      "step": 27170
+    },
+    {
+      "epoch": 0.1788157894736842,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.07607014973958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0248,
+      "loss/crossentropy": 2.2445591568946837,
+      "loss/hidden": 2.8265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23037664219737053,
+      "loss/reg": 0.0,
+      "step": 27180
+    },
+    {
+      "epoch": 0.17888157894736842,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.08280843098958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1128,
+      "loss/crossentropy": 1.8874721288681031,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2291864424943924,
+      "loss/reg": 0.0,
+      "step": 27190
+    },
+    {
+      "epoch": 0.17894736842105263,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.08245035807291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0599,
+      "loss/crossentropy": 2.321315813064575,
+      "loss/hidden": 3.0171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28183609843254087,
+      "loss/reg": 0.0,
+      "step": 27200
+    },
+    {
+      "epoch": 0.17901315789473685,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.04579976399739583,
+      "learning_rate": 0.0001,
+      "loss": 3.1486,
+      "loss/crossentropy": 2.1769141793251037,
+      "loss/hidden": 3.1140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3312571823596954,
+      "loss/reg": 0.0,
+      "step": 27210
+    },
+    {
+      "epoch": 0.17907894736842106,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.0505035400390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0333,
+      "loss/crossentropy": 2.421796774864197,
+      "loss/hidden": 2.6640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2053638830780983,
+      "loss/reg": 0.0,
+      "step": 27220
+    },
+    {
+      "epoch": 0.17914473684210527,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.13408915201822916,
+      "learning_rate": 0.0001,
+      "loss": 3.0911,
+      "loss/crossentropy": 2.539614748954773,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2433784618973732,
+      "loss/reg": 0.0,
+      "step": 27230
+    },
+    {
+      "epoch": 0.1792105263157895,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.08008524576822916,
+      "learning_rate": 0.0001,
+      "loss": 3.0,
+      "loss/crossentropy": 2.1362823486328124,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22769216746091842,
+      "loss/reg": 0.0,
+      "step": 27240
+    },
+    {
+      "epoch": 0.17927631578947367,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.038374837239583334,
+      "learning_rate": 0.0001,
+      "loss": 2.9645,
+      "loss/crossentropy": 2.341468358039856,
+      "loss/hidden": 2.7125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21539552509784698,
+      "loss/reg": 0.0,
+      "step": 27250
+    },
+    {
+      "epoch": 0.17934210526315789,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.031174468994140624,
+      "learning_rate": 0.0001,
+      "loss": 3.0392,
+      "loss/crossentropy": 2.2766498208045958,
+      "loss/hidden": 2.659375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2082691341638565,
+      "loss/reg": 0.0,
+      "step": 27260
+    },
+    {
+      "epoch": 0.1794078947368421,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.16941909790039061,
+      "learning_rate": 0.0001,
+      "loss": 3.0554,
+      "loss/crossentropy": 2.537719178199768,
+      "loss/hidden": 2.9109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21607491597533227,
+      "loss/reg": 0.0,
+      "step": 27270
+    },
+    {
+      "epoch": 0.1794736842105263,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.23950093587239582,
+      "learning_rate": 0.0001,
+      "loss": 2.9943,
+      "loss/crossentropy": 2.4620283365249636,
+      "loss/hidden": 2.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23522212058305741,
+      "loss/reg": 0.0,
+      "step": 27280
+    },
+    {
+      "epoch": 0.17953947368421053,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.1655181884765625,
+      "learning_rate": 0.0001,
+      "loss": 2.9975,
+      "loss/crossentropy": 2.2144778609275817,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22236768230795861,
+      "loss/reg": 0.0,
+      "step": 27290
+    },
+    {
+      "epoch": 0.17960526315789474,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.4901529947916667,
+      "learning_rate": 0.0001,
+      "loss": 3.0091,
+      "loss/crossentropy": 2.1528895676136015,
+      "loss/hidden": 2.69375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2013701803982258,
+      "loss/reg": 0.0,
+      "step": 27300
+    },
+    {
+      "epoch": 0.17967105263157895,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.5030751546223958,
+      "learning_rate": 0.0001,
+      "loss": 2.932,
+      "loss/crossentropy": 2.2105651617050173,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2156506821513176,
+      "loss/reg": 0.0,
+      "step": 27310
+    },
+    {
+      "epoch": 0.17973684210526317,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.1188385009765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0315,
+      "loss/crossentropy": 2.3311607003211976,
+      "loss/hidden": 2.684375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22039272785186767,
+      "loss/reg": 0.0,
+      "step": 27320
+    },
+    {
+      "epoch": 0.17980263157894738,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.10274149576822916,
+      "learning_rate": 0.0001,
+      "loss": 2.9934,
+      "loss/crossentropy": 2.363705575466156,
+      "loss/hidden": 2.9796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2591767907142639,
+      "loss/reg": 0.0,
+      "step": 27330
+    },
+    {
+      "epoch": 0.17986842105263157,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.22634175618489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0232,
+      "loss/crossentropy": 2.491378426551819,
+      "loss/hidden": 2.7171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22361548170447348,
+      "loss/reg": 0.0,
+      "step": 27340
+    },
+    {
+      "epoch": 0.17993421052631578,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.28927408854166664,
+      "learning_rate": 0.0001,
+      "loss": 3.0354,
+      "loss/crossentropy": 2.2730626463890076,
+      "loss/hidden": 2.95,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2728175431489944,
+      "loss/reg": 0.0,
+      "step": 27350
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.2871734619140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0815,
+      "loss/crossentropy": 2.222132349014282,
+      "loss/hidden": 3.0171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27575425505638124,
+      "loss/reg": 0.0,
+      "step": 27360
+    },
+    {
+      "epoch": 0.1800657894736842,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.18010660807291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0514,
+      "loss/crossentropy": 2.3982750535011292,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24716890305280687,
+      "loss/reg": 0.0,
+      "step": 27370
+    },
+    {
+      "epoch": 0.18013157894736842,
+      "grad_norm": 3.625,
+      "grad_norm_var": 0.20855712890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0229,
+      "loss/crossentropy": 2.3283283829689028,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27690982520580293,
+      "loss/reg": 0.0,
+      "step": 27380
+    },
+    {
+      "epoch": 0.18019736842105263,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.14794514973958334,
+      "learning_rate": 0.0001,
+      "loss": 3.1136,
+      "loss/crossentropy": 2.3320053696632383,
+      "loss/hidden": 2.946875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22953213453292848,
+      "loss/reg": 0.0,
+      "step": 27390
+    },
+    {
+      "epoch": 0.18026315789473685,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.15938212076822916,
+      "learning_rate": 0.0001,
+      "loss": 3.0693,
+      "loss/crossentropy": 2.5745165824890135,
+      "loss/hidden": 2.8578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24661683887243271,
+      "loss/reg": 0.0,
+      "step": 27400
+    },
+    {
+      "epoch": 0.18032894736842106,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.08530171712239583,
+      "learning_rate": 0.0001,
+      "loss": 3.027,
+      "loss/crossentropy": 2.291408562660217,
+      "loss/hidden": 2.934375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2568038985133171,
+      "loss/reg": 0.0,
+      "step": 27410
+    },
+    {
+      "epoch": 0.18039473684210527,
+      "grad_norm": 3.0,
+      "grad_norm_var": 0.8119293212890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0758,
+      "loss/crossentropy": 2.275824952125549,
+      "loss/hidden": 2.9875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25695485174655913,
+      "loss/reg": 0.0,
+      "step": 27420
+    },
+    {
+      "epoch": 0.18046052631578946,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.054488118489583334,
+      "learning_rate": 0.0001,
+      "loss": 2.9932,
+      "loss/crossentropy": 2.3114712476730346,
+      "loss/hidden": 2.775,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23609411865472793,
+      "loss/reg": 0.0,
+      "step": 27430
+    },
+    {
+      "epoch": 0.18052631578947367,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.05273030598958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0682,
+      "loss/crossentropy": 2.179815483093262,
+      "loss/hidden": 2.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25313089042901993,
+      "loss/reg": 0.0,
+      "step": 27440
+    },
+    {
+      "epoch": 0.1805921052631579,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.06285807291666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0519,
+      "loss/crossentropy": 2.3046084105968476,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27110700607299804,
+      "loss/reg": 0.0,
+      "step": 27450
+    },
+    {
+      "epoch": 0.1806578947368421,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.06948954264322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0719,
+      "loss/crossentropy": 2.2216897130012514,
+      "loss/hidden": 2.809375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2158276081085205,
+      "loss/reg": 0.0,
+      "step": 27460
+    },
+    {
+      "epoch": 0.18072368421052631,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.06531575520833334,
+      "learning_rate": 0.0001,
+      "loss": 3.0662,
+      "loss/crossentropy": 2.31493022441864,
+      "loss/hidden": 2.86875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2338400349020958,
+      "loss/reg": 0.0,
+      "step": 27470
+    },
+    {
+      "epoch": 0.18078947368421053,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.08967997233072916,
+      "learning_rate": 0.0001,
+      "loss": 3.1017,
+      "loss/crossentropy": 2.292111110687256,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25805368572473525,
+      "loss/reg": 0.0,
+      "step": 27480
+    },
+    {
+      "epoch": 0.18085526315789474,
+      "grad_norm": 3.78125,
+      "grad_norm_var": 0.16054585774739583,
+      "learning_rate": 0.0001,
+      "loss": 3.0896,
+      "loss/crossentropy": 2.531006705760956,
+      "loss/hidden": 2.984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3327594205737114,
+      "loss/reg": 0.0,
+      "step": 27490
+    },
+    {
+      "epoch": 0.18092105263157895,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.1436431884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0498,
+      "loss/crossentropy": 2.3085229277610777,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23342303335666656,
+      "loss/reg": 0.0,
+      "step": 27500
+    },
+    {
+      "epoch": 0.18098684210526317,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.24217122395833332,
+      "learning_rate": 0.0001,
+      "loss": 3.0115,
+      "loss/crossentropy": 2.476685440540314,
+      "loss/hidden": 2.66875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2170659363269806,
+      "loss/reg": 0.0,
+      "step": 27510
+    },
+    {
+      "epoch": 0.18105263157894738,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.13345438639322918,
+      "learning_rate": 0.0001,
+      "loss": 3.0171,
+      "loss/crossentropy": 2.280565071105957,
+      "loss/hidden": 2.6875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20756303817033767,
+      "loss/reg": 0.0,
+      "step": 27520
+    },
+    {
+      "epoch": 0.18111842105263157,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.047200520833333336,
+      "learning_rate": 0.0001,
+      "loss": 2.9912,
+      "loss/crossentropy": 2.553924024105072,
+      "loss/hidden": 2.65625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24719125479459764,
+      "loss/reg": 0.0,
+      "step": 27530
+    },
+    {
+      "epoch": 0.18118421052631578,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.06653238932291666,
+      "learning_rate": 0.0001,
+      "loss": 3.121,
+      "loss/crossentropy": 2.174853026866913,
+      "loss/hidden": 3.0078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31346396207809446,
+      "loss/reg": 0.0,
+      "step": 27540
+    },
+    {
+      "epoch": 0.18125,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.04431966145833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0695,
+      "loss/crossentropy": 2.329943561553955,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22730601876974105,
+      "loss/reg": 0.0,
+      "step": 27550
+    },
+    {
+      "epoch": 0.1813157894736842,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.0733551025390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0813,
+      "loss/crossentropy": 2.2811534643173217,
+      "loss/hidden": 2.9375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24358074963092805,
+      "loss/reg": 0.0,
+      "step": 27560
+    },
+    {
+      "epoch": 0.18138157894736842,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.05774739583333333,
+      "learning_rate": 0.0001,
+      "loss": 3.0169,
+      "loss/crossentropy": 2.3220547437667847,
+      "loss/hidden": 2.70625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24033707976341248,
+      "loss/reg": 0.0,
+      "step": 27570
+    },
+    {
+      "epoch": 0.18144736842105263,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.38362528483072916,
+      "learning_rate": 0.0001,
+      "loss": 3.0383,
+      "loss/crossentropy": 2.4912595987319945,
+      "loss/hidden": 2.809375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2603204816579819,
+      "loss/reg": 0.0,
+      "step": 27580
+    },
+    {
+      "epoch": 0.18151315789473685,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.1091949462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0424,
+      "loss/crossentropy": 2.4132495522499084,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23840930834412574,
+      "loss/reg": 0.0,
+      "step": 27590
+    },
+    {
+      "epoch": 0.18157894736842106,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.224072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0529,
+      "loss/crossentropy": 2.27939612865448,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2168305702507496,
+      "loss/reg": 0.0,
+      "step": 27600
+    },
+    {
+      "epoch": 0.18164473684210528,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.354541015625,
+      "learning_rate": 0.0001,
+      "loss": 3.04,
+      "loss/crossentropy": 2.5797234058380125,
+      "loss/hidden": 2.7140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3075165659189224,
+      "loss/reg": 0.0,
+      "step": 27610
+    },
+    {
+      "epoch": 0.18171052631578946,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.28559544881184895,
+      "learning_rate": 0.0001,
+      "loss": 3.0175,
+      "loss/crossentropy": 2.457747685909271,
+      "loss/hidden": 2.9453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2326775386929512,
+      "loss/reg": 0.0,
+      "step": 27620
+    },
+    {
+      "epoch": 0.18177631578947367,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.2150957743326823,
+      "learning_rate": 0.0001,
+      "loss": 3.025,
+      "loss/crossentropy": 2.2331402122974398,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24202753305435182,
+      "loss/reg": 0.0,
+      "step": 27630
+    },
+    {
+      "epoch": 0.1818421052631579,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.074267578125,
+      "learning_rate": 0.0001,
+      "loss": 3.0184,
+      "loss/crossentropy": 2.402931201457977,
+      "loss/hidden": 2.7140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24029098004102706,
+      "loss/reg": 0.0,
+      "step": 27640
+    },
+    {
+      "epoch": 0.1819078947368421,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.12086181640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0823,
+      "loss/crossentropy": 2.1527361035346986,
+      "loss/hidden": 2.8578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22135660648345948,
+      "loss/reg": 0.0,
+      "step": 27650
+    },
+    {
+      "epoch": 0.18197368421052632,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.13780008951822917,
+      "learning_rate": 0.0001,
+      "loss": 3.0202,
+      "loss/crossentropy": 2.2100749254226684,
+      "loss/hidden": 2.89375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2464701771736145,
+      "loss/reg": 0.0,
+      "step": 27660
+    },
+    {
+      "epoch": 0.18203947368421053,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.08621317545572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0173,
+      "loss/crossentropy": 2.400713062286377,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2229515090584755,
+      "loss/reg": 0.0,
+      "step": 27670
+    },
+    {
+      "epoch": 0.18210526315789474,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.03254292805989583,
+      "learning_rate": 0.0001,
+      "loss": 2.9616,
+      "loss/crossentropy": 2.053451955318451,
+      "loss/hidden": 2.9984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2730902835726738,
+      "loss/reg": 0.0,
+      "step": 27680
+    },
+    {
+      "epoch": 0.18217105263157896,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.05116551717122396,
+      "learning_rate": 0.0001,
+      "loss": 2.9723,
+      "loss/crossentropy": 2.321164917945862,
+      "loss/hidden": 2.6109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2052845761179924,
+      "loss/reg": 0.0,
+      "step": 27690
+    },
+    {
+      "epoch": 0.18223684210526317,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.06518961588541666,
+      "learning_rate": 0.0001,
+      "loss": 3.0474,
+      "loss/crossentropy": 2.1387670636177063,
+      "loss/hidden": 2.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24108488559722902,
+      "loss/reg": 0.0,
+      "step": 27700
+    },
+    {
+      "epoch": 0.18230263157894736,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 4.39171331639651e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.143,
+      "loss/crossentropy": 2.4102607131004334,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33623483330011367,
+      "loss/reg": 0.0,
+      "step": 27710
+    },
+    {
+      "epoch": 0.18236842105263157,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 4.391713315923646e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.053,
+      "loss/crossentropy": 2.215381217002869,
+      "loss/hidden": 2.6921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21635367721319199,
+      "loss/reg": 0.0,
+      "step": 27720
+    },
+    {
+      "epoch": 0.18243421052631578,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.118603515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9848,
+      "loss/crossentropy": 2.456426572799683,
+      "loss/hidden": 2.828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22907530665397643,
+      "loss/reg": 0.0,
+      "step": 27730
+    },
+    {
+      "epoch": 0.1825,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.2470123291015625,
+      "learning_rate": 0.0001,
+      "loss": 2.9824,
+      "loss/crossentropy": 2.0001117050647736,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18826698660850524,
+      "loss/reg": 0.0,
+      "step": 27740
+    },
+    {
+      "epoch": 0.1825657894736842,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.21778971354166668,
+      "learning_rate": 0.0001,
+      "loss": 3.0528,
+      "loss/crossentropy": 2.3025170087814333,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22922251224517823,
+      "loss/reg": 0.0,
+      "step": 27750
+    },
+    {
+      "epoch": 0.18263157894736842,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.07261962890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0966,
+      "loss/crossentropy": 2.299086034297943,
+      "loss/hidden": 2.9109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23557846546173095,
+      "loss/reg": 0.0,
+      "step": 27760
+    },
+    {
+      "epoch": 0.18269736842105264,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.27241109212239584,
+      "learning_rate": 0.0001,
+      "loss": 3.0505,
+      "loss/crossentropy": 2.4849728107452393,
+      "loss/hidden": 2.68125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22220651507377626,
+      "loss/reg": 0.0,
+      "step": 27770
+    },
+    {
+      "epoch": 0.18276315789473685,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.29171549479166664,
+      "learning_rate": 0.0001,
+      "loss": 2.9851,
+      "loss/crossentropy": 2.3156984567642214,
+      "loss/hidden": 2.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.220241579413414,
+      "loss/reg": 0.0,
+      "step": 27780
+    },
+    {
+      "epoch": 0.18282894736842106,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.0520416259765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0197,
+      "loss/crossentropy": 2.4274224162101747,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22342314720153808,
+      "loss/reg": 0.0,
+      "step": 27790
+    },
+    {
+      "epoch": 0.18289473684210528,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.0919342041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0726,
+      "loss/crossentropy": 2.399091196060181,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23315689116716384,
+      "loss/reg": 0.0,
+      "step": 27800
+    },
+    {
+      "epoch": 0.18296052631578946,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.1072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0667,
+      "loss/crossentropy": 2.2596776604652407,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25109176337718964,
+      "loss/reg": 0.0,
+      "step": 27810
+    },
+    {
+      "epoch": 0.18302631578947368,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 1.1102854410807292,
+      "learning_rate": 0.0001,
+      "loss": 3.1493,
+      "loss/crossentropy": 2.2746080636978148,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2303971141576767,
+      "loss/reg": 0.0,
+      "step": 27820
+    },
+    {
+      "epoch": 0.1830921052631579,
+      "grad_norm": 2.25,
+      "grad_norm_var": 1.1916015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0322,
+      "loss/crossentropy": 2.2199900269508364,
+      "loss/hidden": 2.7625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24132085889577864,
+      "loss/reg": 0.0,
+      "step": 27830
+    },
+    {
+      "epoch": 0.1831578947368421,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.11803385416666666,
+      "learning_rate": 0.0001,
+      "loss": 2.9787,
+      "loss/crossentropy": 2.261062300205231,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23909828811883926,
+      "loss/reg": 0.0,
+      "step": 27840
+    },
+    {
+      "epoch": 0.18322368421052632,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.09485270182291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0545,
+      "loss/crossentropy": 2.3090713739395143,
+      "loss/hidden": 2.68125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20118657350540162,
+      "loss/reg": 0.0,
+      "step": 27850
+    },
+    {
+      "epoch": 0.18328947368421053,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.24943008422851562,
+      "learning_rate": 0.0001,
+      "loss": 3.0424,
+      "loss/crossentropy": 2.3824569821357726,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23975446224212646,
+      "loss/reg": 0.0,
+      "step": 27860
+    },
+    {
+      "epoch": 0.18335526315789474,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.14694010416666667,
+      "learning_rate": 0.0001,
+      "loss": 2.9925,
+      "loss/crossentropy": 2.227169597148895,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22001251727342605,
+      "loss/reg": 0.0,
+      "step": 27870
+    },
+    {
+      "epoch": 0.18342105263157896,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.06550191243489584,
+      "learning_rate": 0.0001,
+      "loss": 3.0493,
+      "loss/crossentropy": 2.5089489579200746,
+      "loss/hidden": 2.6625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20695988237857818,
+      "loss/reg": 0.0,
+      "step": 27880
+    },
+    {
+      "epoch": 0.18348684210526317,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.0484527587890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9952,
+      "loss/crossentropy": 2.4806633472442625,
+      "loss/hidden": 2.740625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27523170709609984,
+      "loss/reg": 0.0,
+      "step": 27890
+    },
+    {
+      "epoch": 0.18355263157894736,
+      "grad_norm": 4.3125,
+      "grad_norm_var": 0.2769521077473958,
+      "learning_rate": 0.0001,
+      "loss": 3.0394,
+      "loss/crossentropy": 2.239254057407379,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2308262661099434,
+      "loss/reg": 0.0,
+      "step": 27900
+    },
+    {
+      "epoch": 0.18361842105263157,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.29397761027018227,
+      "learning_rate": 0.0001,
+      "loss": 3.0224,
+      "loss/crossentropy": 2.3714274525642396,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26112145036458967,
+      "loss/reg": 0.0,
+      "step": 27910
+    },
+    {
+      "epoch": 0.18368421052631578,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.05971450805664062,
+      "learning_rate": 0.0001,
+      "loss": 2.9517,
+      "loss/crossentropy": 2.0551604270935058,
+      "loss/hidden": 2.9,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24491064846515656,
+      "loss/reg": 0.0,
+      "step": 27920
+    },
+    {
+      "epoch": 0.18375,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.03795547485351562,
+      "learning_rate": 0.0001,
+      "loss": 3.0193,
+      "loss/crossentropy": 2.3411830067634583,
+      "loss/hidden": 2.8359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23370584547519685,
+      "loss/reg": 0.0,
+      "step": 27930
+    },
+    {
+      "epoch": 0.1838157894736842,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.1084625244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0802,
+      "loss/crossentropy": 2.78390554189682,
+      "loss/hidden": 2.8765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26939452439546585,
+      "loss/reg": 0.0,
+      "step": 27940
+    },
+    {
+      "epoch": 0.18388157894736842,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.41838277180989586,
+      "learning_rate": 0.0001,
+      "loss": 3.1153,
+      "loss/crossentropy": 2.2294405877590178,
+      "loss/hidden": 2.7,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23037993758916855,
+      "loss/reg": 0.0,
+      "step": 27950
+    },
+    {
+      "epoch": 0.18394736842105264,
+      "grad_norm": 3.390625,
+      "grad_norm_var": 0.37854410807291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1324,
+      "loss/crossentropy": 2.28962881565094,
+      "loss/hidden": 2.846875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24506068229675293,
+      "loss/reg": 0.0,
+      "step": 27960
+    },
+    {
+      "epoch": 0.18401315789473685,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.1893999735514323,
+      "learning_rate": 0.0001,
+      "loss": 3.0278,
+      "loss/crossentropy": 2.312084639072418,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2302939549088478,
+      "loss/reg": 0.0,
+      "step": 27970
+    },
+    {
+      "epoch": 0.18407894736842106,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.15084228515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0267,
+      "loss/crossentropy": 2.3484140396118165,
+      "loss/hidden": 2.6046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21540230959653855,
+      "loss/reg": 0.0,
+      "step": 27980
+    },
+    {
+      "epoch": 0.18414473684210525,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.15601806640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0073,
+      "loss/crossentropy": 2.4529793858528137,
+      "loss/hidden": 2.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23430580049753189,
+      "loss/reg": 0.0,
+      "step": 27990
+    },
+    {
+      "epoch": 0.18421052631578946,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.0614166259765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0303,
+      "loss/crossentropy": 2.351777696609497,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2307715743780136,
+      "loss/reg": 0.0,
+      "step": 28000
+    },
+    {
+      "epoch": 0.18427631578947368,
+      "grad_norm": 4.0625,
+      "grad_norm_var": 0.21398824055989582,
+      "learning_rate": 0.0001,
+      "loss": 3.0205,
+      "loss/crossentropy": 2.435424494743347,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2605126053094864,
+      "loss/reg": 0.0,
+      "step": 28010
+    },
+    {
+      "epoch": 0.1843421052631579,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.22447509765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0798,
+      "loss/crossentropy": 2.3683685779571535,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22029414623975754,
+      "loss/reg": 0.0,
+      "step": 28020
+    },
+    {
+      "epoch": 0.1844078947368421,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.08081766764322916,
+      "learning_rate": 0.0001,
+      "loss": 3.0223,
+      "loss/crossentropy": 2.5003761768341066,
+      "loss/hidden": 2.9859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2650137528777122,
+      "loss/reg": 0.0,
+      "step": 28030
+    },
+    {
+      "epoch": 0.18447368421052632,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.07470601399739583,
+      "learning_rate": 0.0001,
+      "loss": 2.998,
+      "loss/crossentropy": 2.2510382771492004,
+      "loss/hidden": 2.7078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23913081735372543,
+      "loss/reg": 0.0,
+      "step": 28040
+    },
+    {
+      "epoch": 0.18453947368421053,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.12652587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0344,
+      "loss/crossentropy": 2.494665837287903,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24169176146388055,
+      "loss/reg": 0.0,
+      "step": 28050
+    },
+    {
+      "epoch": 0.18460526315789474,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.08466389973958334,
+      "learning_rate": 0.0001,
+      "loss": 3.104,
+      "loss/crossentropy": 2.3110733151435854,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22967937737703323,
+      "loss/reg": 0.0,
+      "step": 28060
+    },
+    {
+      "epoch": 0.18467105263157896,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.067431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0115,
+      "loss/crossentropy": 2.4786781549453734,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23725456148386,
+      "loss/reg": 0.0,
+      "step": 28070
+    },
+    {
+      "epoch": 0.18473684210526317,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.07773335774739583,
+      "learning_rate": 0.0001,
+      "loss": 3.0795,
+      "loss/crossentropy": 2.5213263630867004,
+      "loss/hidden": 3.0703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3203597366809845,
+      "loss/reg": 0.0,
+      "step": 28080
+    },
+    {
+      "epoch": 0.18480263157894736,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.06311747233072916,
+      "learning_rate": 0.0001,
+      "loss": 2.9924,
+      "loss/crossentropy": 2.3318725705146788,
+      "loss/hidden": 2.740625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21815032362937928,
+      "loss/reg": 0.0,
+      "step": 28090
+    },
+    {
+      "epoch": 0.18486842105263157,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.42880859375,
+      "learning_rate": 0.0001,
+      "loss": 3.0976,
+      "loss/crossentropy": 2.257493245601654,
+      "loss/hidden": 2.6390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22243027836084367,
+      "loss/reg": 0.0,
+      "step": 28100
+    },
+    {
+      "epoch": 0.18493421052631578,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.16021728515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0607,
+      "loss/crossentropy": 2.320458722114563,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22777790427207947,
+      "loss/reg": 0.0,
+      "step": 28110
+    },
+    {
+      "epoch": 0.185,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.35485738118489585,
+      "learning_rate": 0.0001,
+      "loss": 3.0824,
+      "loss/crossentropy": 2.5934891939163207,
+      "loss/hidden": 2.6625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.217761267721653,
+      "loss/reg": 0.0,
+      "step": 28120
+    },
+    {
+      "epoch": 0.1850657894736842,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.16167577107747397,
+      "learning_rate": 0.0001,
+      "loss": 3.0055,
+      "loss/crossentropy": 2.2394705057144164,
+      "loss/hidden": 2.9890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2515557982027531,
+      "loss/reg": 0.0,
+      "step": 28130
+    },
+    {
+      "epoch": 0.18513157894736842,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.14989598592122397,
+      "learning_rate": 0.0001,
+      "loss": 3.0126,
+      "loss/crossentropy": 2.3399840354919434,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2646253302693367,
+      "loss/reg": 0.0,
+      "step": 28140
+    },
+    {
+      "epoch": 0.18519736842105264,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.14189453125,
+      "learning_rate": 0.0001,
+      "loss": 3.0785,
+      "loss/crossentropy": 2.3720606327056886,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2922904878854752,
+      "loss/reg": 0.0,
+      "step": 28150
+    },
+    {
+      "epoch": 0.18526315789473685,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.052567545572916666,
+      "learning_rate": 0.0001,
+      "loss": 3.0941,
+      "loss/crossentropy": 2.26825897693634,
+      "loss/hidden": 2.7984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24575791507959366,
+      "loss/reg": 0.0,
+      "step": 28160
+    },
+    {
+      "epoch": 0.18532894736842107,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.0409820556640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0183,
+      "loss/crossentropy": 2.2412746131420134,
+      "loss/hidden": 2.7109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22467602118849755,
+      "loss/reg": 0.0,
+      "step": 28170
+    },
+    {
+      "epoch": 0.18539473684210525,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.022264607747395835,
+      "learning_rate": 0.0001,
+      "loss": 3.1005,
+      "loss/crossentropy": 2.465518927574158,
+      "loss/hidden": 2.8625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25958890467882156,
+      "loss/reg": 0.0,
+      "step": 28180
+    },
+    {
+      "epoch": 0.18546052631578946,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.026130167643229167,
+      "learning_rate": 0.0001,
+      "loss": 3.0111,
+      "loss/crossentropy": 2.4506253004074097,
+      "loss/hidden": 2.703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23345130831003189,
+      "loss/reg": 0.0,
+      "step": 28190
+    },
+    {
+      "epoch": 0.18552631578947368,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.2981516520182292,
+      "learning_rate": 0.0001,
+      "loss": 3.1404,
+      "loss/crossentropy": 2.531968724727631,
+      "loss/hidden": 3.134375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3312064751982689,
+      "loss/reg": 0.0,
+      "step": 28200
+    },
+    {
+      "epoch": 0.1855921052631579,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.3452545166015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0669,
+      "loss/crossentropy": 2.1583576440811156,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22225692719221116,
+      "loss/reg": 0.0,
+      "step": 28210
+    },
+    {
+      "epoch": 0.1856578947368421,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.1086090087890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0293,
+      "loss/crossentropy": 2.0994983315467834,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21283998042345048,
+      "loss/reg": 0.0,
+      "step": 28220
+    },
+    {
+      "epoch": 0.18572368421052632,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.09930013020833334,
+      "learning_rate": 0.0001,
+      "loss": 3.0418,
+      "loss/crossentropy": 2.208499777317047,
+      "loss/hidden": 2.6640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21439559012651443,
+      "loss/reg": 0.0,
+      "step": 28230
+    },
+    {
+      "epoch": 0.18578947368421053,
+      "grad_norm": 3.609375,
+      "grad_norm_var": 0.1129302978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1019,
+      "loss/crossentropy": 2.507619249820709,
+      "loss/hidden": 2.9234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2734301760792732,
+      "loss/reg": 0.0,
+      "step": 28240
+    },
+    {
+      "epoch": 0.18585526315789475,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.09721577962239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0004,
+      "loss/crossentropy": 2.440661299228668,
+      "loss/hidden": 2.6890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20598573684692384,
+      "loss/reg": 0.0,
+      "step": 28250
+    },
+    {
+      "epoch": 0.18592105263157896,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.13509012858072916,
+      "learning_rate": 0.0001,
+      "loss": 3.0432,
+      "loss/crossentropy": 2.13590213060379,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21278525814414023,
+      "loss/reg": 0.0,
+      "step": 28260
+    },
+    {
+      "epoch": 0.18598684210526314,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.10868733723958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0082,
+      "loss/crossentropy": 2.4866502404212953,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2882431522011757,
+      "loss/reg": 0.0,
+      "step": 28270
+    },
+    {
+      "epoch": 0.18605263157894736,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.03225504557291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0845,
+      "loss/crossentropy": 1.8784153163433075,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21082937121391296,
+      "loss/reg": 0.0,
+      "step": 28280
+    },
+    {
+      "epoch": 0.18611842105263157,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.0173980712890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0764,
+      "loss/crossentropy": 2.357287549972534,
+      "loss/hidden": 2.646875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20526091530919074,
+      "loss/reg": 0.0,
+      "step": 28290
+    },
+    {
+      "epoch": 0.18618421052631579,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.05768941243489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0812,
+      "loss/crossentropy": 1.8806862443685533,
+      "loss/hidden": 2.621875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18540635257959365,
+      "loss/reg": 0.0,
+      "step": 28300
+    },
+    {
+      "epoch": 0.18625,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.0767730712890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0752,
+      "loss/crossentropy": 2.2633087158203127,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2433292269706726,
+      "loss/reg": 0.0,
+      "step": 28310
+    },
+    {
+      "epoch": 0.1863157894736842,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.05572509765625,
+      "learning_rate": 0.0001,
+      "loss": 2.9722,
+      "loss/crossentropy": 2.328176748752594,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.226905982196331,
+      "loss/reg": 0.0,
+      "step": 28320
+    },
+    {
+      "epoch": 0.18638157894736843,
+      "grad_norm": 1.9765625,
+      "grad_norm_var": 0.08190485636393229,
+      "learning_rate": 0.0001,
+      "loss": 3.0092,
+      "loss/crossentropy": 2.532880795001984,
+      "loss/hidden": 2.971875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3012142822146416,
+      "loss/reg": 0.0,
+      "step": 28330
+    },
+    {
+      "epoch": 0.18644736842105264,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.1636138916015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0108,
+      "loss/crossentropy": 2.823889398574829,
+      "loss/hidden": 2.9046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24069934040308,
+      "loss/reg": 0.0,
+      "step": 28340
+    },
+    {
+      "epoch": 0.18651315789473685,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.10241597493489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0001,
+      "loss/crossentropy": 2.35920227766037,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26395892798900605,
+      "loss/reg": 0.0,
+      "step": 28350
+    },
+    {
+      "epoch": 0.18657894736842107,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.08288472493489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0675,
+      "loss/crossentropy": 2.070932924747467,
+      "loss/hidden": 2.7359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20600728541612626,
+      "loss/reg": 0.0,
+      "step": 28360
+    },
+    {
+      "epoch": 0.18664473684210525,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.07918192545572916,
+      "learning_rate": 0.0001,
+      "loss": 3.017,
+      "loss/crossentropy": 2.307896840572357,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25419241189956665,
+      "loss/reg": 0.0,
+      "step": 28370
+    },
+    {
+      "epoch": 0.18671052631578947,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.05845947265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0325,
+      "loss/crossentropy": 2.482656693458557,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24442674964666367,
+      "loss/reg": 0.0,
+      "step": 28380
+    },
+    {
+      "epoch": 0.18677631578947368,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.12398173014322916,
+      "learning_rate": 0.0001,
+      "loss": 3.1446,
+      "loss/crossentropy": 2.082607638835907,
+      "loss/hidden": 2.975,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2496044546365738,
+      "loss/reg": 0.0,
+      "step": 28390
+    },
+    {
+      "epoch": 0.1868421052631579,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.0723785400390625,
+      "learning_rate": 0.0001,
+      "loss": 3.071,
+      "loss/crossentropy": 2.2511567950248716,
+      "loss/hidden": 2.8765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26172232031822207,
+      "loss/reg": 0.0,
+      "step": 28400
+    },
+    {
+      "epoch": 0.1869078947368421,
+      "grad_norm": 3.09375,
+      "grad_norm_var": 0.08268229166666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1149,
+      "loss/crossentropy": 2.275085437297821,
+      "loss/hidden": 2.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23399723172187806,
+      "loss/reg": 0.0,
+      "step": 28410
+    },
+    {
+      "epoch": 0.18697368421052632,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.1056793212890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0149,
+      "loss/crossentropy": 2.443694305419922,
+      "loss/hidden": 2.8953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26309441328048705,
+      "loss/reg": 0.0,
+      "step": 28420
+    },
+    {
+      "epoch": 0.18703947368421053,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.05745340983072917,
+      "learning_rate": 0.0001,
+      "loss": 3.0672,
+      "loss/crossentropy": 2.1224380493164063,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23011831045150757,
+      "loss/reg": 0.0,
+      "step": 28430
+    },
+    {
+      "epoch": 0.18710526315789475,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.1089752197265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0275,
+      "loss/crossentropy": 2.3803366303443907,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24528718441724778,
+      "loss/reg": 0.0,
+      "step": 28440
+    },
+    {
+      "epoch": 0.18717105263157896,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.1678375244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0534,
+      "loss/crossentropy": 2.254507315158844,
+      "loss/hidden": 2.9875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23182358890771865,
+      "loss/reg": 0.0,
+      "step": 28450
+    },
+    {
+      "epoch": 0.18723684210526315,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.0425201416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0644,
+      "loss/crossentropy": 2.387832987308502,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22640825510025026,
+      "loss/reg": 0.0,
+      "step": 28460
+    },
+    {
+      "epoch": 0.18730263157894736,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.03277587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0762,
+      "loss/crossentropy": 2.4295106649398805,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2517640799283981,
+      "loss/reg": 0.0,
+      "step": 28470
+    },
+    {
+      "epoch": 0.18736842105263157,
+      "grad_norm": 1.96875,
+      "grad_norm_var": 0.15917561848958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0049,
+      "loss/crossentropy": 2.472483456134796,
+      "loss/hidden": 2.6203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21084701418876647,
+      "loss/reg": 0.0,
+      "step": 28480
+    },
+    {
+      "epoch": 0.1874342105263158,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.15578511555989583,
+      "learning_rate": 0.0001,
+      "loss": 3.1143,
+      "loss/crossentropy": 1.9477856278419494,
+      "loss/hidden": 3.04375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26639992743730545,
+      "loss/reg": 0.0,
+      "step": 28490
+    },
+    {
+      "epoch": 0.1875,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.022737630208333335,
+      "learning_rate": 0.0001,
+      "loss": 3.0663,
+      "loss/crossentropy": 2.3596234798431395,
+      "loss/hidden": 2.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27315576672554015,
+      "loss/reg": 0.0,
+      "step": 28500
+    },
+    {
+      "epoch": 0.1875657894736842,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.013753255208333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0261,
+      "loss/crossentropy": 2.196711075305939,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23315538316965104,
+      "loss/reg": 0.0,
+      "step": 28510
+    },
+    {
+      "epoch": 0.18763157894736843,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.05598119099934896,
+      "learning_rate": 0.0001,
+      "loss": 3.1013,
+      "loss/crossentropy": 2.1932213962078095,
+      "loss/hidden": 2.7453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2575216740369797,
+      "loss/reg": 0.0,
+      "step": 28520
+    },
+    {
+      "epoch": 0.18769736842105264,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.0827166239420573,
+      "learning_rate": 0.0001,
+      "loss": 3.0802,
+      "loss/crossentropy": 2.1390640258789064,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21492594629526138,
+      "loss/reg": 0.0,
+      "step": 28530
+    },
+    {
+      "epoch": 0.18776315789473685,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.38186848958333336,
+      "learning_rate": 0.0001,
+      "loss": 3.1353,
+      "loss/crossentropy": 2.3883109092712402,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24425326883792878,
+      "loss/reg": 0.0,
+      "step": 28540
+    },
+    {
+      "epoch": 0.18782894736842104,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.30573628743489584,
+      "learning_rate": 0.0001,
+      "loss": 3.0361,
+      "loss/crossentropy": 2.115387570858002,
+      "loss/hidden": 2.9765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.287911082804203,
+      "loss/reg": 0.0,
+      "step": 28550
+    },
+    {
+      "epoch": 0.18789473684210525,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.07654622395833334,
+      "learning_rate": 0.0001,
+      "loss": 3.1181,
+      "loss/crossentropy": 2.195990490913391,
+      "loss/hidden": 2.840625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22947419285774232,
+      "loss/reg": 0.0,
+      "step": 28560
+    },
+    {
+      "epoch": 0.18796052631578947,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.07486572265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0473,
+      "loss/crossentropy": 2.4482214570045473,
+      "loss/hidden": 2.7234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24045743197202682,
+      "loss/reg": 0.0,
+      "step": 28570
+    },
+    {
+      "epoch": 0.18802631578947368,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.03916600545247396,
+      "learning_rate": 0.0001,
+      "loss": 3.0881,
+      "loss/crossentropy": 2.3246118783950807,
+      "loss/hidden": 2.95625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3818572014570236,
+      "loss/reg": 0.0,
+      "step": 28580
+    },
+    {
+      "epoch": 0.1880921052631579,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.16263427734375,
+      "learning_rate": 0.0001,
+      "loss": 3.0568,
+      "loss/crossentropy": 2.3318132519721986,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25452076345682145,
+      "loss/reg": 0.0,
+      "step": 28590
+    },
+    {
+      "epoch": 0.1881578947368421,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.20039443969726561,
+      "learning_rate": 0.0001,
+      "loss": 3.0826,
+      "loss/crossentropy": 2.0651711583137513,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20308290272951127,
+      "loss/reg": 0.0,
+      "step": 28600
+    },
+    {
+      "epoch": 0.18822368421052632,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.16236750284830728,
+      "learning_rate": 0.0001,
+      "loss": 3.0752,
+      "loss/crossentropy": 2.307542312145233,
+      "loss/hidden": 2.665625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2218876764178276,
+      "loss/reg": 0.0,
+      "step": 28610
+    },
+    {
+      "epoch": 0.18828947368421053,
+      "grad_norm": 4.53125,
+      "grad_norm_var": 0.3535011291503906,
+      "learning_rate": 0.0001,
+      "loss": 3.0657,
+      "loss/crossentropy": 2.2695202469825744,
+      "loss/hidden": 2.8796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24460041224956514,
+      "loss/reg": 0.0,
+      "step": 28620
+    },
+    {
+      "epoch": 0.18835526315789475,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.34776102701822914,
+      "learning_rate": 0.0001,
+      "loss": 3.0931,
+      "loss/crossentropy": 2.409075605869293,
+      "loss/hidden": 3.2140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3033790022134781,
+      "loss/reg": 0.0,
+      "step": 28630
+    },
+    {
+      "epoch": 0.18842105263157893,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.04940999348958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0248,
+      "loss/crossentropy": 2.3460227608680726,
+      "loss/hidden": 2.7390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24943882077932358,
+      "loss/reg": 0.0,
+      "step": 28640
+    },
+    {
+      "epoch": 0.18848684210526315,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.18352762858072916,
+      "learning_rate": 0.0001,
+      "loss": 3.0946,
+      "loss/crossentropy": 2.8061397314071654,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2542914628982544,
+      "loss/reg": 0.0,
+      "step": 28650
+    },
+    {
+      "epoch": 0.18855263157894736,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.05283203125,
+      "learning_rate": 0.0001,
+      "loss": 3.023,
+      "loss/crossentropy": 2.0528534650802612,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2569132924079895,
+      "loss/reg": 0.0,
+      "step": 28660
+    },
+    {
+      "epoch": 0.18861842105263157,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.026520792643229166,
+      "learning_rate": 0.0001,
+      "loss": 2.9614,
+      "loss/crossentropy": 2.2640815138816834,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25332382023334504,
+      "loss/reg": 0.0,
+      "step": 28670
+    },
+    {
+      "epoch": 0.1886842105263158,
+      "grad_norm": 1.953125,
+      "grad_norm_var": 0.037451171875,
+      "learning_rate": 0.0001,
+      "loss": 3.0214,
+      "loss/crossentropy": 2.0984261095523835,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27365772873163224,
+      "loss/reg": 0.0,
+      "step": 28680
+    },
+    {
+      "epoch": 0.18875,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.07590738932291667,
+      "learning_rate": 0.0001,
+      "loss": 3.089,
+      "loss/crossentropy": 2.2889229536056517,
+      "loss/hidden": 2.8453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2945134401321411,
+      "loss/reg": 0.0,
+      "step": 28690
+    },
+    {
+      "epoch": 0.18881578947368421,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.04885152180989583,
+      "learning_rate": 0.0001,
+      "loss": 3.0409,
+      "loss/crossentropy": 2.164911460876465,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22368502318859101,
+      "loss/reg": 0.0,
+      "step": 28700
+    },
+    {
+      "epoch": 0.18888157894736843,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.1095855712890625,
+      "learning_rate": 0.0001,
+      "loss": 3.003,
+      "loss/crossentropy": 2.2275232434272767,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21968200653791428,
+      "loss/reg": 0.0,
+      "step": 28710
+    },
+    {
+      "epoch": 0.18894736842105264,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.05891494750976563,
+      "learning_rate": 0.0001,
+      "loss": 2.985,
+      "loss/crossentropy": 2.429938530921936,
+      "loss/hidden": 2.6046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22446376383304595,
+      "loss/reg": 0.0,
+      "step": 28720
+    },
+    {
+      "epoch": 0.18901315789473686,
+      "grad_norm": 1.984375,
+      "grad_norm_var": 0.13444722493489583,
+      "learning_rate": 0.0001,
+      "loss": 2.9926,
+      "loss/crossentropy": 2.139791202545166,
+      "loss/hidden": 2.8578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2599222779273987,
+      "loss/reg": 0.0,
+      "step": 28730
+    },
+    {
+      "epoch": 0.18907894736842104,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.15026041666666667,
+      "learning_rate": 0.0001,
+      "loss": 2.9502,
+      "loss/crossentropy": 2.5147767782211305,
+      "loss/hidden": 2.609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20266549810767173,
+      "loss/reg": 0.0,
+      "step": 28740
+    },
+    {
+      "epoch": 0.18914473684210525,
+      "grad_norm": 3.578125,
+      "grad_norm_var": 0.20552978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.051,
+      "loss/crossentropy": 2.4063800454139708,
+      "loss/hidden": 3.08125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3112260654568672,
+      "loss/reg": 0.0,
+      "step": 28750
+    },
+    {
+      "epoch": 0.18921052631578947,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.14830322265625,
+      "learning_rate": 0.0001,
+      "loss": 3.057,
+      "loss/crossentropy": 2.2508232951164246,
+      "loss/hidden": 2.725,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22451825439929962,
+      "loss/reg": 0.0,
+      "step": 28760
+    },
+    {
+      "epoch": 0.18927631578947368,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.11416727701822917,
+      "learning_rate": 0.0001,
+      "loss": 3.0512,
+      "loss/crossentropy": 2.149078315496445,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21534078270196916,
+      "loss/reg": 0.0,
+      "step": 28770
+    },
+    {
+      "epoch": 0.1893421052631579,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.61968994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0034,
+      "loss/crossentropy": 2.3696099877357484,
+      "loss/hidden": 2.715625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2401156485080719,
+      "loss/reg": 0.0,
+      "step": 28780
+    },
+    {
+      "epoch": 0.1894078947368421,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.4911092122395833,
+      "learning_rate": 0.0001,
+      "loss": 3.0951,
+      "loss/crossentropy": 2.0147340178489683,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2364984579384327,
+      "loss/reg": 0.0,
+      "step": 28790
+    },
+    {
+      "epoch": 0.18947368421052632,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.06545817057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0985,
+      "loss/crossentropy": 2.4977503657341003,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25097098499536513,
+      "loss/reg": 0.0,
+      "step": 28800
+    },
+    {
+      "epoch": 0.18953947368421054,
+      "grad_norm": 1.96875,
+      "grad_norm_var": 0.08662007649739584,
+      "learning_rate": 0.0001,
+      "loss": 3.0047,
+      "loss/crossentropy": 2.1395092368125916,
+      "loss/hidden": 2.740625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20759201645851136,
+      "loss/reg": 0.0,
+      "step": 28810
+    },
+    {
+      "epoch": 0.18960526315789475,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.09783426920572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0329,
+      "loss/crossentropy": 2.3047179579734802,
+      "loss/hidden": 2.765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22515993937849998,
+      "loss/reg": 0.0,
+      "step": 28820
+    },
+    {
+      "epoch": 0.18967105263157893,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.07600504557291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0786,
+      "loss/crossentropy": 2.4169238924980165,
+      "loss/hidden": 2.765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23725131005048752,
+      "loss/reg": 0.0,
+      "step": 28830
+    },
+    {
+      "epoch": 0.18973684210526315,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.09014460245768229,
+      "learning_rate": 0.0001,
+      "loss": 3.0935,
+      "loss/crossentropy": 2.473812985420227,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23891154676675797,
+      "loss/reg": 0.0,
+      "step": 28840
+    },
+    {
+      "epoch": 0.18980263157894736,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.3475870768229167,
+      "learning_rate": 0.0001,
+      "loss": 3.0779,
+      "loss/crossentropy": 2.1757280230522156,
+      "loss/hidden": 2.9328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2538435861468315,
+      "loss/reg": 0.0,
+      "step": 28850
+    },
+    {
+      "epoch": 0.18986842105263158,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.2500233968098958,
+      "learning_rate": 0.0001,
+      "loss": 3.0282,
+      "loss/crossentropy": 2.321690630912781,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2390053778886795,
+      "loss/reg": 0.0,
+      "step": 28860
+    },
+    {
+      "epoch": 0.1899342105263158,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.12604878743489584,
+      "learning_rate": 0.0001,
+      "loss": 3.057,
+      "loss/crossentropy": 2.0470433115959166,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24679146856069564,
+      "loss/reg": 0.0,
+      "step": 28870
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.17348531087239583,
+      "learning_rate": 0.0001,
+      "loss": 3.1491,
+      "loss/crossentropy": 2.022314542531967,
+      "loss/hidden": 2.86875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23445970118045806,
+      "loss/reg": 0.0,
+      "step": 28880
+    },
+    {
+      "epoch": 0.19006578947368422,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.06368815104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0043,
+      "loss/crossentropy": 2.0953749775886537,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23144432604312898,
+      "loss/reg": 0.0,
+      "step": 28890
+    },
+    {
+      "epoch": 0.19013157894736843,
+      "grad_norm": 2.0,
+      "grad_norm_var": 0.04041239420572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0328,
+      "loss/crossentropy": 2.2653061032295225,
+      "loss/hidden": 2.7578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23457860052585602,
+      "loss/reg": 0.0,
+      "step": 28900
+    },
+    {
+      "epoch": 0.19019736842105264,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.04980061848958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0439,
+      "loss/crossentropy": 2.222883141040802,
+      "loss/hidden": 3.0453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2797882482409477,
+      "loss/reg": 0.0,
+      "step": 28910
+    },
+    {
+      "epoch": 0.19026315789473683,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.0666168212890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0709,
+      "loss/crossentropy": 2.3731595158576964,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21689550429582596,
+      "loss/reg": 0.0,
+      "step": 28920
+    },
+    {
+      "epoch": 0.19032894736842104,
+      "grad_norm": 3.5,
+      "grad_norm_var": 0.15308837890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0906,
+      "loss/crossentropy": 2.3481295704841614,
+      "loss/hidden": 2.5953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20085408240556718,
+      "loss/reg": 0.0,
+      "step": 28930
+    },
+    {
+      "epoch": 0.19039473684210526,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.427294667561849,
+      "learning_rate": 0.0001,
+      "loss": 3.0422,
+      "loss/crossentropy": 2.1098085761070253,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20854987427592278,
+      "loss/reg": 0.0,
+      "step": 28940
+    },
+    {
+      "epoch": 0.19046052631578947,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 2.635267893473307,
+      "learning_rate": 0.0001,
+      "loss": 3.0915,
+      "loss/crossentropy": 2.1917474389076235,
+      "loss/hidden": 2.7203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20776809751987457,
+      "loss/reg": 0.0,
+      "step": 28950
+    },
+    {
+      "epoch": 0.19052631578947368,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.062235514322916664,
+      "learning_rate": 0.0001,
+      "loss": 3.0466,
+      "loss/crossentropy": 2.0871485114097594,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21217636987566948,
+      "loss/reg": 0.0,
+      "step": 28960
+    },
+    {
+      "epoch": 0.1905921052631579,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.028055826822916668,
+      "learning_rate": 0.0001,
+      "loss": 3.0856,
+      "loss/crossentropy": 2.1593764424324036,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20974511429667472,
+      "loss/reg": 0.0,
+      "step": 28970
+    },
+    {
+      "epoch": 0.1906578947368421,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.09622395833333333,
+      "learning_rate": 0.0001,
+      "loss": 2.983,
+      "loss/crossentropy": 2.2977413177490233,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25048716068267823,
+      "loss/reg": 0.0,
+      "step": 28980
+    },
+    {
+      "epoch": 0.19072368421052632,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.11470947265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0439,
+      "loss/crossentropy": 2.235085117816925,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2309124991297722,
+      "loss/reg": 0.0,
+      "step": 28990
+    },
+    {
+      "epoch": 0.19078947368421054,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.23596598307291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0173,
+      "loss/crossentropy": 2.5475085377693176,
+      "loss/hidden": 2.7265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25446673631668093,
+      "loss/reg": 0.0,
+      "step": 29000
+    },
+    {
+      "epoch": 0.19085526315789475,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.18063151041666667,
+      "learning_rate": 0.0001,
+      "loss": 3.0599,
+      "loss/crossentropy": 2.327496898174286,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2621834442019463,
+      "loss/reg": 0.0,
+      "step": 29010
+    },
+    {
+      "epoch": 0.19092105263157894,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.0766754150390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0494,
+      "loss/crossentropy": 2.0966503024101257,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21290977895259858,
+      "loss/reg": 0.0,
+      "step": 29020
+    },
+    {
+      "epoch": 0.19098684210526315,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.0940338134765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0697,
+      "loss/crossentropy": 2.193461000919342,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2083764299750328,
+      "loss/reg": 0.0,
+      "step": 29030
+    },
+    {
+      "epoch": 0.19105263157894736,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.20330785115559896,
+      "learning_rate": 0.0001,
+      "loss": 3.0073,
+      "loss/crossentropy": 2.1486512422561646,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2170367144048214,
+      "loss/reg": 0.0,
+      "step": 29040
+    },
+    {
+      "epoch": 0.19111842105263158,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.05134989420572917,
+      "learning_rate": 0.0001,
+      "loss": 2.9996,
+      "loss/crossentropy": 2.359115946292877,
+      "loss/hidden": 2.7390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2235071614384651,
+      "loss/reg": 0.0,
+      "step": 29050
+    },
+    {
+      "epoch": 0.1911842105263158,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.05135498046875,
+      "learning_rate": 0.0001,
+      "loss": 3.0619,
+      "loss/crossentropy": 2.4227041006088257,
+      "loss/hidden": 2.978125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.316775880753994,
+      "loss/reg": 0.0,
+      "step": 29060
+    },
+    {
+      "epoch": 0.19125,
+      "grad_norm": 3.375,
+      "grad_norm_var": 0.10562235514322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0286,
+      "loss/crossentropy": 2.532416009902954,
+      "loss/hidden": 2.69375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2569827824831009,
+      "loss/reg": 0.0,
+      "step": 29070
+    },
+    {
+      "epoch": 0.19131578947368422,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.11773173014322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0518,
+      "loss/crossentropy": 2.35415917634964,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23699928000569342,
+      "loss/reg": 0.0,
+      "step": 29080
+    },
+    {
+      "epoch": 0.19138157894736843,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.03673502604166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0043,
+      "loss/crossentropy": 2.044126057624817,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23696889132261276,
+      "loss/reg": 0.0,
+      "step": 29090
+    },
+    {
+      "epoch": 0.19144736842105264,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.08551432291666666,
+      "learning_rate": 0.0001,
+      "loss": 3.1223,
+      "loss/crossentropy": 2.4426622867584227,
+      "loss/hidden": 2.8859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.251794858276844,
+      "loss/reg": 0.0,
+      "step": 29100
+    },
+    {
+      "epoch": 0.19151315789473683,
+      "grad_norm": 1.9921875,
+      "grad_norm_var": 0.2218523661295573,
+      "learning_rate": 0.0001,
+      "loss": 3.069,
+      "loss/crossentropy": 2.5686187982559203,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.250452946126461,
+      "loss/reg": 0.0,
+      "step": 29110
+    },
+    {
+      "epoch": 0.19157894736842104,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.4294288635253906,
+      "learning_rate": 0.0001,
+      "loss": 3.0377,
+      "loss/crossentropy": 2.2918658018112184,
+      "loss/hidden": 2.596875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21649599373340606,
+      "loss/reg": 0.0,
+      "step": 29120
+    },
+    {
+      "epoch": 0.19164473684210526,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.028401692708333332,
+      "learning_rate": 0.0001,
+      "loss": 3.1086,
+      "loss/crossentropy": 2.3098790526390074,
+      "loss/hidden": 2.7375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2144581601023674,
+      "loss/reg": 0.0,
+      "step": 29130
+    },
+    {
+      "epoch": 0.19171052631578947,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.020048014322916665,
+      "learning_rate": 0.0001,
+      "loss": 3.0103,
+      "loss/crossentropy": 2.470157337188721,
+      "loss/hidden": 2.8875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2766988605260849,
+      "loss/reg": 0.0,
+      "step": 29140
+    },
+    {
+      "epoch": 0.19177631578947368,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.02623291015625,
+      "learning_rate": 0.0001,
+      "loss": 2.9856,
+      "loss/crossentropy": 2.317296016216278,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2372336909174919,
+      "loss/reg": 0.0,
+      "step": 29150
+    },
+    {
+      "epoch": 0.1918421052631579,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.0629547119140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0697,
+      "loss/crossentropy": 2.1767319798469544,
+      "loss/hidden": 3.0015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2674897871911526,
+      "loss/reg": 0.0,
+      "step": 29160
+    },
+    {
+      "epoch": 0.1919078947368421,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.05247294108072917,
+      "learning_rate": 0.0001,
+      "loss": 2.9929,
+      "loss/crossentropy": 2.2281210064888,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24070803225040435,
+      "loss/reg": 0.0,
+      "step": 29170
+    },
+    {
+      "epoch": 0.19197368421052632,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.30436909993489586,
+      "learning_rate": 0.0001,
+      "loss": 3.0128,
+      "loss/crossentropy": 2.2079304993152618,
+      "loss/hidden": 2.7453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20276689082384108,
+      "loss/reg": 0.0,
+      "step": 29180
+    },
+    {
+      "epoch": 0.19203947368421054,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.331884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0248,
+      "loss/crossentropy": 2.3135082483291627,
+      "loss/hidden": 2.7,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2161063551902771,
+      "loss/reg": 0.0,
+      "step": 29190
+    },
+    {
+      "epoch": 0.19210526315789472,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.15064697265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0372,
+      "loss/crossentropy": 2.377930212020874,
+      "loss/hidden": 2.7234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22175273448228836,
+      "loss/reg": 0.0,
+      "step": 29200
+    },
+    {
+      "epoch": 0.19217105263157894,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.12327372233072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1012,
+      "loss/crossentropy": 2.4531391739845274,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2535667777061462,
+      "loss/reg": 0.0,
+      "step": 29210
+    },
+    {
+      "epoch": 0.19223684210526315,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.07448628743489584,
+      "learning_rate": 0.0001,
+      "loss": 3.1458,
+      "loss/crossentropy": 2.2929787397384644,
+      "loss/hidden": 2.996875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25961560308933257,
+      "loss/reg": 0.0,
+      "step": 29220
+    },
+    {
+      "epoch": 0.19230263157894736,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.04540608723958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0156,
+      "loss/crossentropy": 2.265605902671814,
+      "loss/hidden": 2.953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27446902841329573,
+      "loss/reg": 0.0,
+      "step": 29230
+    },
+    {
+      "epoch": 0.19236842105263158,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.06565348307291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1152,
+      "loss/crossentropy": 2.376231300830841,
+      "loss/hidden": 2.890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2864221647381783,
+      "loss/reg": 0.0,
+      "step": 29240
+    },
+    {
+      "epoch": 0.1924342105263158,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.06280085245768229,
+      "learning_rate": 0.0001,
+      "loss": 3.07,
+      "loss/crossentropy": 2.394097375869751,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2377302184700966,
+      "loss/reg": 0.0,
+      "step": 29250
+    },
+    {
+      "epoch": 0.1925,
+      "grad_norm": 3.40625,
+      "grad_norm_var": 0.122607421875,
+      "learning_rate": 0.0001,
+      "loss": 3.1114,
+      "loss/crossentropy": 2.5634241580963133,
+      "loss/hidden": 2.7140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22698014378547668,
+      "loss/reg": 0.0,
+      "step": 29260
+    },
+    {
+      "epoch": 0.19256578947368422,
+      "grad_norm": 3.171875,
+      "grad_norm_var": 0.11873372395833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0884,
+      "loss/crossentropy": 2.3539575576782226,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2572378695011139,
+      "loss/reg": 0.0,
+      "step": 29270
+    },
+    {
+      "epoch": 0.19263157894736843,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.14241536458333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0603,
+      "loss/crossentropy": 2.3354499697685243,
+      "loss/hidden": 3.003125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2519584596157074,
+      "loss/reg": 0.0,
+      "step": 29280
+    },
+    {
+      "epoch": 0.19269736842105264,
+      "grad_norm": 3.5625,
+      "grad_norm_var": 0.3069081624348958,
+      "learning_rate": 0.0001,
+      "loss": 3.0894,
+      "loss/crossentropy": 2.0824127376079558,
+      "loss/hidden": 2.89375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28106222823262217,
+      "loss/reg": 0.0,
+      "step": 29290
+    },
+    {
+      "epoch": 0.19276315789473683,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.1501617431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0121,
+      "loss/crossentropy": 2.1776723742485045,
+      "loss/hidden": 2.6359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20980774164199828,
+      "loss/reg": 0.0,
+      "step": 29300
+    },
+    {
+      "epoch": 0.19282894736842104,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.2937164306640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1225,
+      "loss/crossentropy": 2.157558262348175,
+      "loss/hidden": 3.0953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2775505542755127,
+      "loss/reg": 0.0,
+      "step": 29310
+    },
+    {
+      "epoch": 0.19289473684210526,
+      "grad_norm": 3.734375,
+      "grad_norm_var": 0.33560791015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0748,
+      "loss/crossentropy": 2.381651961803436,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2246159717440605,
+      "loss/reg": 0.0,
+      "step": 29320
+    },
+    {
+      "epoch": 0.19296052631578947,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 2.1095540364583334,
+      "learning_rate": 0.0001,
+      "loss": 3.2928,
+      "loss/crossentropy": 2.0944801807403564,
+      "loss/hidden": 2.7375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23710142374038695,
+      "loss/reg": 0.0,
+      "step": 29330
+    },
+    {
+      "epoch": 0.19302631578947368,
+      "grad_norm": 1.9453125,
+      "grad_norm_var": 0.2770851135253906,
+      "learning_rate": 0.0001,
+      "loss": 3.1,
+      "loss/crossentropy": 2.4732042074203493,
+      "loss/hidden": 2.7515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23788287490606308,
+      "loss/reg": 0.0,
+      "step": 29340
+    },
+    {
+      "epoch": 0.1930921052631579,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.35339330037434896,
+      "learning_rate": 0.0001,
+      "loss": 3.0744,
+      "loss/crossentropy": 2.4158215165138244,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23991909474134446,
+      "loss/reg": 0.0,
+      "step": 29350
+    },
+    {
+      "epoch": 0.1931578947368421,
+      "grad_norm": 4.03125,
+      "grad_norm_var": 0.3200754801432292,
+      "learning_rate": 0.0001,
+      "loss": 3.1003,
+      "loss/crossentropy": 2.448209798336029,
+      "loss/hidden": 3.3078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3389866009354591,
+      "loss/reg": 0.0,
+      "step": 29360
+    },
+    {
+      "epoch": 0.19322368421052633,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.22730712890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1074,
+      "loss/crossentropy": 1.9778976082801818,
+      "loss/hidden": 3.0078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2316925495862961,
+      "loss/reg": 0.0,
+      "step": 29370
+    },
+    {
+      "epoch": 0.19328947368421054,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.09205729166666667,
+      "learning_rate": 0.0001,
+      "loss": 3.1044,
+      "loss/crossentropy": 2.344563841819763,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2164514496922493,
+      "loss/reg": 0.0,
+      "step": 29380
+    },
+    {
+      "epoch": 0.19335526315789472,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.07641499837239583,
+      "learning_rate": 0.0001,
+      "loss": 3.1445,
+      "loss/crossentropy": 2.186362612247467,
+      "loss/hidden": 2.7203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21316134929656982,
+      "loss/reg": 0.0,
+      "step": 29390
+    },
+    {
+      "epoch": 0.19342105263157894,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.5996897379557292,
+      "learning_rate": 0.0001,
+      "loss": 3.11,
+      "loss/crossentropy": 2.546473169326782,
+      "loss/hidden": 2.70625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22721548825502397,
+      "loss/reg": 0.0,
+      "step": 29400
+    },
+    {
+      "epoch": 0.19348684210526315,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.6576881408691406,
+      "learning_rate": 0.0001,
+      "loss": 2.9865,
+      "loss/crossentropy": 2.210515594482422,
+      "loss/hidden": 2.6765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20338982120156288,
+      "loss/reg": 0.0,
+      "step": 29410
+    },
+    {
+      "epoch": 0.19355263157894737,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.06096572875976562,
+      "learning_rate": 0.0001,
+      "loss": 3.0997,
+      "loss/crossentropy": 2.5028133630752563,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3376652091741562,
+      "loss/reg": 0.0,
+      "step": 29420
+    },
+    {
+      "epoch": 0.19361842105263158,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.25368626912434894,
+      "learning_rate": 0.0001,
+      "loss": 3.0177,
+      "loss/crossentropy": 2.2334850907325743,
+      "loss/hidden": 2.99375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2385883465409279,
+      "loss/reg": 0.0,
+      "step": 29430
+    },
+    {
+      "epoch": 0.1936842105263158,
+      "grad_norm": 1.9765625,
+      "grad_norm_var": 0.24181493123372397,
+      "learning_rate": 0.0001,
+      "loss": 3.0405,
+      "loss/crossentropy": 1.898663866519928,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20240384489297866,
+      "loss/reg": 0.0,
+      "step": 29440
+    },
+    {
+      "epoch": 0.19375,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.03530044555664062,
+      "learning_rate": 0.0001,
+      "loss": 3.0841,
+      "loss/crossentropy": 2.3499651670455934,
+      "loss/hidden": 2.93125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27134826183319094,
+      "loss/reg": 0.0,
+      "step": 29450
+    },
+    {
+      "epoch": 0.19381578947368422,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.02574462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0395,
+      "loss/crossentropy": 2.3818777322769167,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25071706622838974,
+      "loss/reg": 0.0,
+      "step": 29460
+    },
+    {
+      "epoch": 0.19388157894736843,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.018602498372395835,
+      "learning_rate": 0.0001,
+      "loss": 3.0232,
+      "loss/crossentropy": 2.1610496282577514,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2543530076742172,
+      "loss/reg": 0.0,
+      "step": 29470
+    },
+    {
+      "epoch": 0.19394736842105262,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.12858784993489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0346,
+      "loss/crossentropy": 2.3739108026027678,
+      "loss/hidden": 2.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22633943632245063,
+      "loss/reg": 0.0,
+      "step": 29480
+    },
+    {
+      "epoch": 0.19401315789473683,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.116162109375,
+      "learning_rate": 0.0001,
+      "loss": 3.0797,
+      "loss/crossentropy": 2.3364730775356293,
+      "loss/hidden": 2.7421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2235547423362732,
+      "loss/reg": 0.0,
+      "step": 29490
+    },
+    {
+      "epoch": 0.19407894736842105,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.04695612589518229,
+      "learning_rate": 0.0001,
+      "loss": 3.0498,
+      "loss/crossentropy": 2.2981515765190124,
+      "loss/hidden": 2.9703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3249427303671837,
+      "loss/reg": 0.0,
+      "step": 29500
+    },
+    {
+      "epoch": 0.19414473684210526,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.06395848592122395,
+      "learning_rate": 0.0001,
+      "loss": 3.0062,
+      "loss/crossentropy": 2.392876994609833,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2336449593305588,
+      "loss/reg": 0.0,
+      "step": 29510
+    },
+    {
+      "epoch": 0.19421052631578947,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.07955322265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0426,
+      "loss/crossentropy": 2.3927242517471314,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23498377203941345,
+      "loss/reg": 0.0,
+      "step": 29520
+    },
+    {
+      "epoch": 0.19427631578947369,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.030354817708333332,
+      "learning_rate": 0.0001,
+      "loss": 3.0498,
+      "loss/crossentropy": 1.9698969006538392,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2563909277319908,
+      "loss/reg": 0.0,
+      "step": 29530
+    },
+    {
+      "epoch": 0.1943421052631579,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.08201395670572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0928,
+      "loss/crossentropy": 2.228722929954529,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2367407873272896,
+      "loss/reg": 0.0,
+      "step": 29540
+    },
+    {
+      "epoch": 0.1944078947368421,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.067626953125,
+      "learning_rate": 0.0001,
+      "loss": 3.0422,
+      "loss/crossentropy": 2.4088454604148866,
+      "loss/hidden": 2.8140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24344853162765503,
+      "loss/reg": 0.0,
+      "step": 29550
+    },
+    {
+      "epoch": 0.19447368421052633,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.07860921223958334,
+      "learning_rate": 0.0001,
+      "loss": 3.06,
+      "loss/crossentropy": 2.1968480169773104,
+      "loss/hidden": 2.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22112407311797141,
+      "loss/reg": 0.0,
+      "step": 29560
+    },
+    {
+      "epoch": 0.19453947368421054,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.2692942301432292,
+      "learning_rate": 0.0001,
+      "loss": 3.0565,
+      "loss/crossentropy": 2.3560967564582826,
+      "loss/hidden": 3.040625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2932572916150093,
+      "loss/reg": 0.0,
+      "step": 29570
+    },
+    {
+      "epoch": 0.19460526315789473,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.30523859659830727,
+      "learning_rate": 0.0001,
+      "loss": 3.0003,
+      "loss/crossentropy": 2.0988304018974304,
+      "loss/hidden": 2.8859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24731282144784927,
+      "loss/reg": 0.0,
+      "step": 29580
+    },
+    {
+      "epoch": 0.19467105263157894,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.055944569905598956,
+      "learning_rate": 0.0001,
+      "loss": 3.0501,
+      "loss/crossentropy": 2.2525598287582396,
+      "loss/hidden": 2.671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20050241351127623,
+      "loss/reg": 0.0,
+      "step": 29590
+    },
+    {
+      "epoch": 0.19473684210526315,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.038557942708333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0499,
+      "loss/crossentropy": 2.117692744731903,
+      "loss/hidden": 2.7140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20816649496555328,
+      "loss/reg": 0.0,
+      "step": 29600
+    },
+    {
+      "epoch": 0.19480263157894737,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.048628743489583334,
+      "learning_rate": 0.0001,
+      "loss": 3.0671,
+      "loss/crossentropy": 2.2347262859344483,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24547438323497772,
+      "loss/reg": 0.0,
+      "step": 29610
+    },
+    {
+      "epoch": 0.19486842105263158,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.1297027587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1136,
+      "loss/crossentropy": 2.1894510865211485,
+      "loss/hidden": 2.9796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26874606013298036,
+      "loss/reg": 0.0,
+      "step": 29620
+    },
+    {
+      "epoch": 0.1949342105263158,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.11825764973958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1032,
+      "loss/crossentropy": 2.314576745033264,
+      "loss/hidden": 2.9359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2493584305047989,
+      "loss/reg": 0.0,
+      "step": 29630
+    },
+    {
+      "epoch": 0.195,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.18583577473958332,
+      "learning_rate": 0.0001,
+      "loss": 3.0667,
+      "loss/crossentropy": 2.548297035694122,
+      "loss/hidden": 2.69375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24562474042177201,
+      "loss/reg": 0.0,
+      "step": 29640
+    },
+    {
+      "epoch": 0.19506578947368422,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.21096903483072918,
+      "learning_rate": 0.0001,
+      "loss": 3.0832,
+      "loss/crossentropy": 2.685001492500305,
+      "loss/hidden": 2.8703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27619747072458267,
+      "loss/reg": 0.0,
+      "step": 29650
+    },
+    {
+      "epoch": 0.19513157894736843,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.10445556640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0406,
+      "loss/crossentropy": 2.471345865726471,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.233747598528862,
+      "loss/reg": 0.0,
+      "step": 29660
+    },
+    {
+      "epoch": 0.19519736842105262,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.1791412353515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0572,
+      "loss/crossentropy": 2.378851294517517,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.222721017152071,
+      "loss/reg": 0.0,
+      "step": 29670
+    },
+    {
+      "epoch": 0.19526315789473683,
+      "grad_norm": 3.328125,
+      "grad_norm_var": 0.25057144165039064,
+      "learning_rate": 0.0001,
+      "loss": 3.027,
+      "loss/crossentropy": 2.091909795999527,
+      "loss/hidden": 2.8578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24436787366867066,
+      "loss/reg": 0.0,
+      "step": 29680
+    },
+    {
+      "epoch": 0.19532894736842105,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.22474339803059895,
+      "learning_rate": 0.0001,
+      "loss": 3.0429,
+      "loss/crossentropy": 2.328868269920349,
+      "loss/hidden": 2.7453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.262276391685009,
+      "loss/reg": 0.0,
+      "step": 29690
+    },
+    {
+      "epoch": 0.19539473684210526,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.15204671223958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0549,
+      "loss/crossentropy": 2.296106255054474,
+      "loss/hidden": 2.9734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2577256761491299,
+      "loss/reg": 0.0,
+      "step": 29700
+    },
+    {
+      "epoch": 0.19546052631578947,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.20373433430989582,
+      "learning_rate": 0.0001,
+      "loss": 3.0571,
+      "loss/crossentropy": 2.495807719230652,
+      "loss/hidden": 2.90625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2599473804235458,
+      "loss/reg": 0.0,
+      "step": 29710
+    },
+    {
+      "epoch": 0.1955263157894737,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.73043212890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0451,
+      "loss/crossentropy": 2.489153337478638,
+      "loss/hidden": 2.7421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2395859479904175,
+      "loss/reg": 0.0,
+      "step": 29720
+    },
+    {
+      "epoch": 0.1955921052631579,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.09890034993489584,
+      "learning_rate": 0.0001,
+      "loss": 2.9846,
+      "loss/crossentropy": 2.1876869797706604,
+      "loss/hidden": 2.684375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20331611335277558,
+      "loss/reg": 0.0,
+      "step": 29730
+    },
+    {
+      "epoch": 0.19565789473684211,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.12862040201822916,
+      "learning_rate": 0.0001,
+      "loss": 3.0944,
+      "loss/crossentropy": 2.548052740097046,
+      "loss/hidden": 2.6890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23612861186265946,
+      "loss/reg": 0.0,
+      "step": 29740
+    },
+    {
+      "epoch": 0.19572368421052633,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.038802083333333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0041,
+      "loss/crossentropy": 2.450440764427185,
+      "loss/hidden": 2.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20733811408281327,
+      "loss/reg": 0.0,
+      "step": 29750
+    },
+    {
+      "epoch": 0.1957894736842105,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.035497029622395836,
+      "learning_rate": 0.0001,
+      "loss": 3.0924,
+      "loss/crossentropy": 2.022493052482605,
+      "loss/hidden": 3.0265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26215304881334306,
+      "loss/reg": 0.0,
+      "step": 29760
+    },
+    {
+      "epoch": 0.19585526315789473,
+      "grad_norm": 6.1875,
+      "grad_norm_var": 0.9495025634765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0461,
+      "loss/crossentropy": 2.2956669092178346,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24869376718997954,
+      "loss/reg": 0.0,
+      "step": 29770
+    },
+    {
+      "epoch": 0.19592105263157894,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 1.1040974934895833,
+      "learning_rate": 0.0001,
+      "loss": 3.0038,
+      "loss/crossentropy": 2.301337730884552,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2174853652715683,
+      "loss/reg": 0.0,
+      "step": 29780
+    },
+    {
+      "epoch": 0.19598684210526315,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.26876627604166664,
+      "learning_rate": 0.0001,
+      "loss": 3.1862,
+      "loss/crossentropy": 1.7189460813999176,
+      "loss/hidden": 2.953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21047319620847701,
+      "loss/reg": 0.0,
+      "step": 29790
+    },
+    {
+      "epoch": 0.19605263157894737,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.042464192708333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0449,
+      "loss/crossentropy": 2.3193385720252992,
+      "loss/hidden": 2.9546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2548700511455536,
+      "loss/reg": 0.0,
+      "step": 29800
+    },
+    {
+      "epoch": 0.19611842105263158,
+      "grad_norm": 3.71875,
+      "grad_norm_var": 0.27180887858072916,
+      "learning_rate": 0.0001,
+      "loss": 3.008,
+      "loss/crossentropy": 2.3117380261421205,
+      "loss/hidden": 2.646875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2046105980873108,
+      "loss/reg": 0.0,
+      "step": 29810
+    },
+    {
+      "epoch": 0.1961842105263158,
+      "grad_norm": 4.28125,
+      "grad_norm_var": 0.516106923421224,
+      "learning_rate": 0.0001,
+      "loss": 3.0925,
+      "loss/crossentropy": 2.3800249338150024,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26439079344272615,
+      "loss/reg": 0.0,
+      "step": 29820
+    },
+    {
+      "epoch": 0.19625,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.3475870768229167,
+      "learning_rate": 0.0001,
+      "loss": 3.0886,
+      "loss/crossentropy": 2.284978838264942,
+      "loss/hidden": 2.9421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24723461624234916,
+      "loss/reg": 0.0,
+      "step": 29830
+    },
+    {
+      "epoch": 0.19631578947368422,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.29708836873372396,
+      "learning_rate": 0.0001,
+      "loss": 3.0244,
+      "loss/crossentropy": 2.231731951236725,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3149738535284996,
+      "loss/reg": 0.0,
+      "step": 29840
+    },
+    {
+      "epoch": 0.19638157894736843,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.287847646077474,
+      "learning_rate": 0.0001,
+      "loss": 3.0728,
+      "loss/crossentropy": 2.453169012069702,
+      "loss/hidden": 2.809375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25355182588100433,
+      "loss/reg": 0.0,
+      "step": 29850
+    },
+    {
+      "epoch": 0.19644736842105262,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.2615386962890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0352,
+      "loss/crossentropy": 2.174148201942444,
+      "loss/hidden": 2.66875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2032485894858837,
+      "loss/reg": 0.0,
+      "step": 29860
+    },
+    {
+      "epoch": 0.19651315789473683,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.25632705688476565,
+      "learning_rate": 0.0001,
+      "loss": 3.0585,
+      "loss/crossentropy": 2.483177053928375,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2651928335428238,
+      "loss/reg": 0.0,
+      "step": 29870
+    },
+    {
+      "epoch": 0.19657894736842105,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.12946370442708333,
+      "learning_rate": 0.0001,
+      "loss": 3.0076,
+      "loss/crossentropy": 2.4216135859489443,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2323567435145378,
+      "loss/reg": 0.0,
+      "step": 29880
+    },
+    {
+      "epoch": 0.19664473684210526,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.07423477172851563,
+      "learning_rate": 0.0001,
+      "loss": 2.9992,
+      "loss/crossentropy": 2.43337767124176,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2374792516231537,
+      "loss/reg": 0.0,
+      "step": 29890
+    },
+    {
+      "epoch": 0.19671052631578947,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.09332275390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1552,
+      "loss/crossentropy": 2.066901612281799,
+      "loss/hidden": 2.9546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2512317180633545,
+      "loss/reg": 0.0,
+      "step": 29900
+    },
+    {
+      "epoch": 0.1967763157894737,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.3540598551432292,
+      "learning_rate": 0.0001,
+      "loss": 3.0271,
+      "loss/crossentropy": 2.4260810136795046,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26536442786455156,
+      "loss/reg": 0.0,
+      "step": 29910
+    },
+    {
+      "epoch": 0.1968421052631579,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.10589090983072917,
+      "learning_rate": 0.0001,
+      "loss": 3.0213,
+      "loss/crossentropy": 2.4159180164337157,
+      "loss/hidden": 2.6640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2190088465809822,
+      "loss/reg": 0.0,
+      "step": 29920
+    },
+    {
+      "epoch": 0.19690789473684212,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.03914286295572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0847,
+      "loss/crossentropy": 2.061979150772095,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20891676545143129,
+      "loss/reg": 0.0,
+      "step": 29930
+    },
+    {
+      "epoch": 0.19697368421052633,
+      "grad_norm": 3.625,
+      "grad_norm_var": 0.14216206868489584,
+      "learning_rate": 0.0001,
+      "loss": 3.0243,
+      "loss/crossentropy": 2.06211262345314,
+      "loss/hidden": 2.8390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22453068271279336,
+      "loss/reg": 0.0,
+      "step": 29940
+    },
+    {
+      "epoch": 0.19703947368421051,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.13738505045572916,
+      "learning_rate": 0.0001,
+      "loss": 2.9814,
+      "loss/crossentropy": 2.1103883236646652,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1844348356127739,
+      "loss/reg": 0.0,
+      "step": 29950
+    },
+    {
+      "epoch": 0.19710526315789473,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.026537068684895835,
+      "learning_rate": 0.0001,
+      "loss": 3.0008,
+      "loss/crossentropy": 2.553321826457977,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21997541338205337,
+      "loss/reg": 0.0,
+      "step": 29960
+    },
+    {
+      "epoch": 0.19717105263157894,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.02939453125,
+      "learning_rate": 0.0001,
+      "loss": 3.0382,
+      "loss/crossentropy": 2.4657467365264893,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2679098337888718,
+      "loss/reg": 0.0,
+      "step": 29970
+    },
+    {
+      "epoch": 0.19723684210526315,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.32203369140625,
+      "learning_rate": 0.0001,
+      "loss": 3.083,
+      "loss/crossentropy": 2.2925270318984987,
+      "loss/hidden": 3.0421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2816790759563446,
+      "loss/reg": 0.0,
+      "step": 29980
+    },
+    {
+      "epoch": 0.19730263157894737,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.3014149983723958,
+      "learning_rate": 0.0001,
+      "loss": 3.0358,
+      "loss/crossentropy": 2.652257299423218,
+      "loss/hidden": 2.665625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23255081921815873,
+      "loss/reg": 0.0,
+      "step": 29990
+    },
+    {
+      "epoch": 0.19736842105263158,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.061799112955729166,
+      "learning_rate": 0.0001,
+      "loss": 2.9585,
+      "loss/crossentropy": 2.3400336265563966,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2366102933883667,
+      "loss/reg": 0.0,
+      "step": 30000
+    },
+    {
+      "epoch": 0.1974342105263158,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.06998672485351562,
+      "learning_rate": 0.0001,
+      "loss": 3.0787,
+      "loss/crossentropy": 2.2949488759040833,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24081108272075652,
+      "loss/reg": 0.0,
+      "step": 30010
+    },
+    {
+      "epoch": 0.1975,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.41442769368489585,
+      "learning_rate": 0.0001,
+      "loss": 3.0288,
+      "loss/crossentropy": 2.414297103881836,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23484020233154296,
+      "loss/reg": 0.0,
+      "step": 30020
+    },
+    {
+      "epoch": 0.19756578947368422,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.18356704711914062,
+      "learning_rate": 0.0001,
+      "loss": 3.0326,
+      "loss/crossentropy": 2.300756549835205,
+      "loss/hidden": 2.5453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19037204384803771,
+      "loss/reg": 0.0,
+      "step": 30030
+    },
+    {
+      "epoch": 0.1976315789473684,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.07932942708333333,
+      "learning_rate": 0.0001,
+      "loss": 3.0541,
+      "loss/crossentropy": 2.5434207677841187,
+      "loss/hidden": 2.6859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2503844425082207,
+      "loss/reg": 0.0,
+      "step": 30040
+    },
+    {
+      "epoch": 0.19769736842105262,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.03378499348958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0416,
+      "loss/crossentropy": 2.435779368877411,
+      "loss/hidden": 2.6671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22367140799760818,
+      "loss/reg": 0.0,
+      "step": 30050
+    },
+    {
+      "epoch": 0.19776315789473684,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.11531575520833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0175,
+      "loss/crossentropy": 2.614890933036804,
+      "loss/hidden": 2.7125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2448316603899002,
+      "loss/reg": 0.0,
+      "step": 30060
+    },
+    {
+      "epoch": 0.19782894736842105,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.13720703125,
+      "learning_rate": 0.0001,
+      "loss": 3.0335,
+      "loss/crossentropy": 2.0392467260360716,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22009356170892716,
+      "loss/reg": 0.0,
+      "step": 30070
+    },
+    {
+      "epoch": 0.19789473684210526,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.028351847330729166,
+      "learning_rate": 0.0001,
+      "loss": 3.0004,
+      "loss/crossentropy": 2.463162088394165,
+      "loss/hidden": 2.6109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2216602995991707,
+      "loss/reg": 0.0,
+      "step": 30080
+    },
+    {
+      "epoch": 0.19796052631578948,
+      "grad_norm": 1.859375,
+      "grad_norm_var": 0.05168863932291667,
+      "learning_rate": 0.0001,
+      "loss": 2.9309,
+      "loss/crossentropy": 2.386566638946533,
+      "loss/hidden": 2.665625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23080737441778182,
+      "loss/reg": 0.0,
+      "step": 30090
+    },
+    {
+      "epoch": 0.1980263157894737,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.40987955729166664,
+      "learning_rate": 0.0001,
+      "loss": 2.9946,
+      "loss/crossentropy": 2.5444513320922852,
+      "loss/hidden": 2.6296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22146540880203247,
+      "loss/reg": 0.0,
+      "step": 30100
+    },
+    {
+      "epoch": 0.1980921052631579,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.45288263956705727,
+      "learning_rate": 0.0001,
+      "loss": 3.0193,
+      "loss/crossentropy": 2.485434365272522,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23654820621013642,
+      "loss/reg": 0.0,
+      "step": 30110
+    },
+    {
+      "epoch": 0.19815789473684212,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.20391820271809896,
+      "learning_rate": 0.0001,
+      "loss": 3.0528,
+      "loss/crossentropy": 2.378788614273071,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2376804992556572,
+      "loss/reg": 0.0,
+      "step": 30120
+    },
+    {
+      "epoch": 0.1982236842105263,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.10331624348958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0188,
+      "loss/crossentropy": 2.154638743400574,
+      "loss/hidden": 2.6875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20261103957891463,
+      "loss/reg": 0.0,
+      "step": 30130
+    },
+    {
+      "epoch": 0.19828947368421052,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.08589655558268229,
+      "learning_rate": 0.0001,
+      "loss": 2.9379,
+      "loss/crossentropy": 2.2816161513328552,
+      "loss/hidden": 2.6921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19446674287319182,
+      "loss/reg": 0.0,
+      "step": 30140
+    },
+    {
+      "epoch": 0.19835526315789473,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.10558980305989583,
+      "learning_rate": 0.0001,
+      "loss": 3.0039,
+      "loss/crossentropy": 2.333029532432556,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21476658284664155,
+      "loss/reg": 0.0,
+      "step": 30150
+    },
+    {
+      "epoch": 0.19842105263157894,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.08321940104166667,
+      "learning_rate": 0.0001,
+      "loss": 2.9808,
+      "loss/crossentropy": 2.155623471736908,
+      "loss/hidden": 2.5859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18124678283929824,
+      "loss/reg": 0.0,
+      "step": 30160
+    },
+    {
+      "epoch": 0.19848684210526316,
+      "grad_norm": 1.9453125,
+      "grad_norm_var": 0.13827311197916667,
+      "learning_rate": 0.0001,
+      "loss": 3.0116,
+      "loss/crossentropy": 2.109735357761383,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24903104603290557,
+      "loss/reg": 0.0,
+      "step": 30170
+    },
+    {
+      "epoch": 0.19855263157894737,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.040135701497395836,
+      "learning_rate": 0.0001,
+      "loss": 3.0087,
+      "loss/crossentropy": 2.290113925933838,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21611811220645905,
+      "loss/reg": 0.0,
+      "step": 30180
+    },
+    {
+      "epoch": 0.19861842105263158,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.04804280598958333,
+      "learning_rate": 0.0001,
+      "loss": 2.995,
+      "loss/crossentropy": 1.9855021834373474,
+      "loss/hidden": 2.6984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1928846351802349,
+      "loss/reg": 0.0,
+      "step": 30190
+    },
+    {
+      "epoch": 0.1986842105263158,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.059366607666015626,
+      "learning_rate": 0.0001,
+      "loss": 3.0265,
+      "loss/crossentropy": 2.213775265216827,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2045104242861271,
+      "loss/reg": 0.0,
+      "step": 30200
+    },
+    {
+      "epoch": 0.19875,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.10831883748372396,
+      "learning_rate": 0.0001,
+      "loss": 3.0971,
+      "loss/crossentropy": 2.2141775846481324,
+      "loss/hidden": 2.9078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2511264935135841,
+      "loss/reg": 0.0,
+      "step": 30210
+    },
+    {
+      "epoch": 0.19881578947368422,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.18625895182291666,
+      "learning_rate": 0.0001,
+      "loss": 3.1376,
+      "loss/crossentropy": 2.201507192850113,
+      "loss/hidden": 3.01875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27276814319193365,
+      "loss/reg": 0.0,
+      "step": 30220
+    },
+    {
+      "epoch": 0.1988815789473684,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.1700347900390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0325,
+      "loss/crossentropy": 2.4042665481567385,
+      "loss/hidden": 2.78125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22706928849220276,
+      "loss/reg": 0.0,
+      "step": 30230
+    },
+    {
+      "epoch": 0.19894736842105262,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.18758036295572916,
+      "learning_rate": 0.0001,
+      "loss": 3.049,
+      "loss/crossentropy": 2.0507681727409364,
+      "loss/hidden": 2.946875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22476800307631492,
+      "loss/reg": 0.0,
+      "step": 30240
+    },
+    {
+      "epoch": 0.19901315789473684,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.07414957682291666,
+      "learning_rate": 0.0001,
+      "loss": 2.9881,
+      "loss/crossentropy": 2.2702227234840393,
+      "loss/hidden": 2.69375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2327321708202362,
+      "loss/reg": 0.0,
+      "step": 30250
+    },
+    {
+      "epoch": 0.19907894736842105,
+      "grad_norm": 3.40625,
+      "grad_norm_var": 0.29836832682291664,
+      "learning_rate": 0.0001,
+      "loss": 3.1337,
+      "loss/crossentropy": 2.3777806520462037,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2500802963972092,
+      "loss/reg": 0.0,
+      "step": 30260
+    },
+    {
+      "epoch": 0.19914473684210526,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.27034403483072916,
+      "learning_rate": 0.0001,
+      "loss": 3.0123,
+      "loss/crossentropy": 2.431059980392456,
+      "loss/hidden": 2.965625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2794636771082878,
+      "loss/reg": 0.0,
+      "step": 30270
+    },
+    {
+      "epoch": 0.19921052631578948,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.022591145833333333,
+      "learning_rate": 0.0001,
+      "loss": 3.0564,
+      "loss/crossentropy": 2.491440224647522,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27763516157865525,
+      "loss/reg": 0.0,
+      "step": 30280
+    },
+    {
+      "epoch": 0.1992763157894737,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.24989827473958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0591,
+      "loss/crossentropy": 2.443405735492706,
+      "loss/hidden": 2.8,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2293305829167366,
+      "loss/reg": 0.0,
+      "step": 30290
+    },
+    {
+      "epoch": 0.1993421052631579,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.044205729166666666,
+      "learning_rate": 0.0001,
+      "loss": 2.9622,
+      "loss/crossentropy": 2.2449229061603546,
+      "loss/hidden": 2.734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24314892143011094,
+      "loss/reg": 0.0,
+      "step": 30300
+    },
+    {
+      "epoch": 0.19940789473684212,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.03411458333333333,
+      "learning_rate": 0.0001,
+      "loss": 3.038,
+      "loss/crossentropy": 2.5837315797805784,
+      "loss/hidden": 2.846875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25667293965816496,
+      "loss/reg": 0.0,
+      "step": 30310
+    },
+    {
+      "epoch": 0.1994736842105263,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.03482666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.03,
+      "loss/crossentropy": 2.4695034623146057,
+      "loss/hidden": 2.6765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24289859235286712,
+      "loss/reg": 0.0,
+      "step": 30320
+    },
+    {
+      "epoch": 0.19953947368421052,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.08709309895833334,
+      "learning_rate": 0.0001,
+      "loss": 2.9926,
+      "loss/crossentropy": 2.263628613948822,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25994745194911956,
+      "loss/reg": 0.0,
+      "step": 30330
+    },
+    {
+      "epoch": 0.19960526315789473,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.12483317057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0437,
+      "loss/crossentropy": 2.263992178440094,
+      "loss/hidden": 2.8234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22506246864795684,
+      "loss/reg": 0.0,
+      "step": 30340
+    },
+    {
+      "epoch": 0.19967105263157894,
+      "grad_norm": 2.0,
+      "grad_norm_var": 0.09973042805989583,
+      "learning_rate": 0.0001,
+      "loss": 3.0212,
+      "loss/crossentropy": 2.4082891941070557,
+      "loss/hidden": 2.6890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24035572707653047,
+      "loss/reg": 0.0,
+      "step": 30350
+    },
+    {
+      "epoch": 0.19973684210526316,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.04556884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0589,
+      "loss/crossentropy": 2.2902461647987367,
+      "loss/hidden": 2.671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20010970458388327,
+      "loss/reg": 0.0,
+      "step": 30360
+    },
+    {
+      "epoch": 0.19980263157894737,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.047098795572916664,
+      "learning_rate": 0.0001,
+      "loss": 3.0519,
+      "loss/crossentropy": 2.260884428024292,
+      "loss/hidden": 2.9109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23876612037420272,
+      "loss/reg": 0.0,
+      "step": 30370
+    },
+    {
+      "epoch": 0.19986842105263158,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.17197265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0596,
+      "loss/crossentropy": 2.293976974487305,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24686284512281417,
+      "loss/reg": 0.0,
+      "step": 30380
+    },
+    {
+      "epoch": 0.1999342105263158,
+      "grad_norm": 1.96875,
+      "grad_norm_var": 0.4962565104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0143,
+      "loss/crossentropy": 2.0594027996063233,
+      "loss/hidden": 2.8625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2534930631518364,
+      "loss/reg": 0.0,
+      "step": 30390
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.5117421468098958,
+      "learning_rate": 0.0001,
+      "loss": 3.0243,
+      "loss/crossentropy": 2.2039811968803407,
+      "loss/hidden": 2.8765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2378777429461479,
+      "loss/reg": 0.0,
+      "step": 30400
+    },
+    {
+      "epoch": 0.2000657894736842,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.06318257649739584,
+      "learning_rate": 0.0001,
+      "loss": 3.0318,
+      "loss/crossentropy": 2.306100535392761,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25150387436151506,
+      "loss/reg": 0.0,
+      "step": 30410
+    },
+    {
+      "epoch": 0.2001315789473684,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.048249308268229166,
+      "learning_rate": 0.0001,
+      "loss": 3.0054,
+      "loss/crossentropy": 2.154050374031067,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23610089272260665,
+      "loss/reg": 0.0,
+      "step": 30420
+    },
+    {
+      "epoch": 0.20019736842105262,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.1066314697265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0415,
+      "loss/crossentropy": 2.3580272018909456,
+      "loss/hidden": 2.7375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21100984290242195,
+      "loss/reg": 0.0,
+      "step": 30430
+    },
+    {
+      "epoch": 0.20026315789473684,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.10287984212239583,
+      "learning_rate": 0.0001,
+      "loss": 3.021,
+      "loss/crossentropy": 2.3812718391418457,
+      "loss/hidden": 2.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2449337661266327,
+      "loss/reg": 0.0,
+      "step": 30440
+    },
+    {
+      "epoch": 0.20032894736842105,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.05493876139322917,
+      "learning_rate": 0.0001,
+      "loss": 2.9937,
+      "loss/crossentropy": 2.406654155254364,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28666631430387496,
+      "loss/reg": 0.0,
+      "step": 30450
+    },
+    {
+      "epoch": 0.20039473684210526,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.0645904541015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1228,
+      "loss/crossentropy": 2.3998040676116945,
+      "loss/hidden": 2.9015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2500763192772865,
+      "loss/reg": 0.0,
+      "step": 30460
+    },
+    {
+      "epoch": 0.20046052631578948,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.1094146728515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9741,
+      "loss/crossentropy": 2.1604729771614073,
+      "loss/hidden": 2.590625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19615808725357056,
+      "loss/reg": 0.0,
+      "step": 30470
+    },
+    {
+      "epoch": 0.2005263157894737,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.1017974853515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0351,
+      "loss/crossentropy": 2.3099865555763244,
+      "loss/hidden": 2.78125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2315102458000183,
+      "loss/reg": 0.0,
+      "step": 30480
+    },
+    {
+      "epoch": 0.2005921052631579,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.41324462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0974,
+      "loss/crossentropy": 2.3289413452148438,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2756644278764725,
+      "loss/reg": 0.0,
+      "step": 30490
+    },
+    {
+      "epoch": 0.20065789473684212,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 1.2159505208333334,
+      "learning_rate": 0.0001,
+      "loss": 2.9922,
+      "loss/crossentropy": 2.2126728296279907,
+      "loss/hidden": 2.8203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22418673560023308,
+      "loss/reg": 0.0,
+      "step": 30500
+    },
+    {
+      "epoch": 0.2007236842105263,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.9500343322753906,
+      "learning_rate": 0.0001,
+      "loss": 2.9726,
+      "loss/crossentropy": 2.3763804376125335,
+      "loss/hidden": 2.7578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22033569663763047,
+      "loss/reg": 0.0,
+      "step": 30510
+    },
+    {
+      "epoch": 0.20078947368421052,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.14224853515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0223,
+      "loss/crossentropy": 2.400152790546417,
+      "loss/hidden": 2.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21929119899868965,
+      "loss/reg": 0.0,
+      "step": 30520
+    },
+    {
+      "epoch": 0.20085526315789473,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 103.514990234375,
+      "learning_rate": 0.0001,
+      "loss": 3.0504,
+      "loss/crossentropy": 2.1772961974143983,
+      "loss/hidden": 3.1234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23286587446928025,
+      "loss/reg": 0.0,
+      "step": 30530
+    },
+    {
+      "epoch": 0.20092105263157894,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.11510391235351562,
+      "learning_rate": 0.0001,
+      "loss": 3.0667,
+      "loss/crossentropy": 2.2763838887214662,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22003234028816224,
+      "loss/reg": 0.0,
+      "step": 30540
+    },
+    {
+      "epoch": 0.20098684210526316,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.08960367838541666,
+      "learning_rate": 0.0001,
+      "loss": 2.9552,
+      "loss/crossentropy": 2.418367159366608,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2519043207168579,
+      "loss/reg": 0.0,
+      "step": 30550
+    },
+    {
+      "epoch": 0.20105263157894737,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.07086588541666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0293,
+      "loss/crossentropy": 2.4948137521743776,
+      "loss/hidden": 2.871875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24468859434127807,
+      "loss/reg": 0.0,
+      "step": 30560
+    },
+    {
+      "epoch": 0.20111842105263159,
+      "grad_norm": 4.21875,
+      "grad_norm_var": 0.2363677978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0781,
+      "loss/crossentropy": 2.697233200073242,
+      "loss/hidden": 2.9953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2648474559187889,
+      "loss/reg": 0.0,
+      "step": 30570
+    },
+    {
+      "epoch": 0.2011842105263158,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.26945699055989586,
+      "learning_rate": 0.0001,
+      "loss": 2.9756,
+      "loss/crossentropy": 2.3618945240974427,
+      "loss/hidden": 2.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2500054851174355,
+      "loss/reg": 0.0,
+      "step": 30580
+    },
+    {
+      "epoch": 0.20125,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.09013264973958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0325,
+      "loss/crossentropy": 2.1298818975687026,
+      "loss/hidden": 2.5125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1715396959334612,
+      "loss/reg": 0.0,
+      "step": 30590
+    },
+    {
+      "epoch": 0.2013157894736842,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.09729715983072916,
+      "learning_rate": 0.0001,
+      "loss": 3.0119,
+      "loss/crossentropy": 2.204944038391113,
+      "loss/hidden": 2.790625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21538405269384384,
+      "loss/reg": 0.0,
+      "step": 30600
+    },
+    {
+      "epoch": 0.2013815789473684,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.03004150390625,
+      "learning_rate": 0.0001,
+      "loss": 2.9775,
+      "loss/crossentropy": 2.4482433080673216,
+      "loss/hidden": 2.6703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22656193226575852,
+      "loss/reg": 0.0,
+      "step": 30610
+    },
+    {
+      "epoch": 0.20144736842105262,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.03572769165039062,
+      "learning_rate": 0.0001,
+      "loss": 3.0182,
+      "loss/crossentropy": 2.0842928767204283,
+      "loss/hidden": 2.6734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2059495523571968,
+      "loss/reg": 0.0,
+      "step": 30620
+    },
+    {
+      "epoch": 0.20151315789473684,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.1525286356608073,
+      "learning_rate": 0.0001,
+      "loss": 3.0799,
+      "loss/crossentropy": 2.3205880761146545,
+      "loss/hidden": 2.8265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22585390508174896,
+      "loss/reg": 0.0,
+      "step": 30630
+    },
+    {
+      "epoch": 0.20157894736842105,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.08941650390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0334,
+      "loss/crossentropy": 2.287918508052826,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23107146471738815,
+      "loss/reg": 0.0,
+      "step": 30640
+    },
+    {
+      "epoch": 0.20164473684210527,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.0238433837890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9747,
+      "loss/crossentropy": 2.4189619421958923,
+      "loss/hidden": 2.653125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2165643572807312,
+      "loss/reg": 0.0,
+      "step": 30650
+    },
+    {
+      "epoch": 0.20171052631578948,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.07734349568684896,
+      "learning_rate": 0.0001,
+      "loss": 2.9809,
+      "loss/crossentropy": 2.484665501117706,
+      "loss/hidden": 2.815625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25413639694452284,
+      "loss/reg": 0.0,
+      "step": 30660
+    },
+    {
+      "epoch": 0.2017763157894737,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.13760477701822918,
+      "learning_rate": 0.0001,
+      "loss": 3.0598,
+      "loss/crossentropy": 1.993758463859558,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23019451200962066,
+      "loss/reg": 0.0,
+      "step": 30670
+    },
+    {
+      "epoch": 0.2018421052631579,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.16373697916666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0066,
+      "loss/crossentropy": 2.508138656616211,
+      "loss/hidden": 2.740625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29257247895002364,
+      "loss/reg": 0.0,
+      "step": 30680
+    },
+    {
+      "epoch": 0.2019078947368421,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.16415786743164062,
+      "learning_rate": 0.0001,
+      "loss": 3.0694,
+      "loss/crossentropy": 2.178850865364075,
+      "loss/hidden": 2.6546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19227326661348343,
+      "loss/reg": 0.0,
+      "step": 30690
+    },
+    {
+      "epoch": 0.2019736842105263,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.12350234985351563,
+      "learning_rate": 0.0001,
+      "loss": 3.0484,
+      "loss/crossentropy": 2.395913541316986,
+      "loss/hidden": 2.7203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23005425035953522,
+      "loss/reg": 0.0,
+      "step": 30700
+    },
+    {
+      "epoch": 0.20203947368421052,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.07646077473958333,
+      "learning_rate": 0.0001,
+      "loss": 2.9845,
+      "loss/crossentropy": 2.0765319585800173,
+      "loss/hidden": 2.828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21056767106056212,
+      "loss/reg": 0.0,
+      "step": 30710
+    },
+    {
+      "epoch": 0.20210526315789473,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.14062093098958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0582,
+      "loss/crossentropy": 2.1704251885414125,
+      "loss/hidden": 2.7296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.217359322309494,
+      "loss/reg": 0.0,
+      "step": 30720
+    },
+    {
+      "epoch": 0.20217105263157895,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.1090484619140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0285,
+      "loss/crossentropy": 2.1888681292533874,
+      "loss/hidden": 2.7140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23789106756448747,
+      "loss/reg": 0.0,
+      "step": 30730
+    },
+    {
+      "epoch": 0.20223684210526316,
+      "grad_norm": 1.9921875,
+      "grad_norm_var": 0.2569536844889323,
+      "learning_rate": 0.0001,
+      "loss": 3.0386,
+      "loss/crossentropy": 2.161084806919098,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23299970030784606,
+      "loss/reg": 0.0,
+      "step": 30740
+    },
+    {
+      "epoch": 0.20230263157894737,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.14261245727539062,
+      "learning_rate": 0.0001,
+      "loss": 2.9869,
+      "loss/crossentropy": 2.4508134722709656,
+      "loss/hidden": 2.6078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20573023706674576,
+      "loss/reg": 0.0,
+      "step": 30750
+    },
+    {
+      "epoch": 0.2023684210526316,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.08206761678059896,
+      "learning_rate": 0.0001,
+      "loss": 3.1039,
+      "loss/crossentropy": 2.5278658986091616,
+      "loss/hidden": 2.909375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2716485261917114,
+      "loss/reg": 0.0,
+      "step": 30760
+    },
+    {
+      "epoch": 0.2024342105263158,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.14897435506184895,
+      "learning_rate": 0.0001,
+      "loss": 3.0202,
+      "loss/crossentropy": 2.2195838689804077,
+      "loss/hidden": 2.9484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29928734600543977,
+      "loss/reg": 0.0,
+      "step": 30770
+    },
+    {
+      "epoch": 0.2025,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.3571248372395833,
+      "learning_rate": 0.0001,
+      "loss": 2.9956,
+      "loss/crossentropy": 1.9182006061077117,
+      "loss/hidden": 2.840625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2259790524840355,
+      "loss/reg": 0.0,
+      "step": 30780
+    },
+    {
+      "epoch": 0.2025657894736842,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.3158365885416667,
+      "learning_rate": 0.0001,
+      "loss": 3.043,
+      "loss/crossentropy": 2.2211042761802675,
+      "loss/hidden": 2.690625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23300761282444,
+      "loss/reg": 0.0,
+      "step": 30790
+    },
+    {
+      "epoch": 0.2026315789473684,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.1349609375,
+      "learning_rate": 0.0001,
+      "loss": 3.046,
+      "loss/crossentropy": 2.2689759850502016,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22895248532295226,
+      "loss/reg": 0.0,
+      "step": 30800
+    },
+    {
+      "epoch": 0.20269736842105263,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.03440653483072917,
+      "learning_rate": 0.0001,
+      "loss": 2.9736,
+      "loss/crossentropy": 2.2549991250038146,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24080493301153183,
+      "loss/reg": 0.0,
+      "step": 30810
+    },
+    {
+      "epoch": 0.20276315789473684,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 1.0274088541666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0971,
+      "loss/crossentropy": 2.3225981116294863,
+      "loss/hidden": 2.9296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24680722504854202,
+      "loss/reg": 0.0,
+      "step": 30820
+    },
+    {
+      "epoch": 0.20282894736842105,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 1.2762102762858072,
+      "learning_rate": 0.0001,
+      "loss": 3.0638,
+      "loss/crossentropy": 2.355189287662506,
+      "loss/hidden": 2.775,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23843924850225448,
+      "loss/reg": 0.0,
+      "step": 30830
+    },
+    {
+      "epoch": 0.20289473684210527,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 2.2499407450358073,
+      "learning_rate": 0.0001,
+      "loss": 3.2035,
+      "loss/crossentropy": 2.2234743475914,
+      "loss/hidden": 2.934375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2502759709954262,
+      "loss/reg": 0.0,
+      "step": 30840
+    },
+    {
+      "epoch": 0.20296052631578948,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 2.0923665364583335,
+      "learning_rate": 0.0001,
+      "loss": 3.0222,
+      "loss/crossentropy": 2.535164365172386,
+      "loss/hidden": 2.7078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21734880432486534,
+      "loss/reg": 0.0,
+      "step": 30850
+    },
+    {
+      "epoch": 0.2030263157894737,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.03389383951822917,
+      "learning_rate": 0.0001,
+      "loss": 2.9959,
+      "loss/crossentropy": 2.4880054354667664,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2341625601053238,
+      "loss/reg": 0.0,
+      "step": 30860
+    },
+    {
+      "epoch": 0.2030921052631579,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.0282135009765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0451,
+      "loss/crossentropy": 2.277066648006439,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2593472898006439,
+      "loss/reg": 0.0,
+      "step": 30870
+    },
+    {
+      "epoch": 0.2031578947368421,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.03958333333333333,
+      "learning_rate": 0.0001,
+      "loss": 3.0158,
+      "loss/crossentropy": 2.1649017184972763,
+      "loss/hidden": 2.7125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2148670382797718,
+      "loss/reg": 0.0,
+      "step": 30880
+    },
+    {
+      "epoch": 0.2032236842105263,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.37971903483072916,
+      "learning_rate": 0.0001,
+      "loss": 2.9899,
+      "loss/crossentropy": 2.3688385248184205,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2734518602490425,
+      "loss/reg": 0.0,
+      "step": 30890
+    },
+    {
+      "epoch": 0.20328947368421052,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.3890452067057292,
+      "learning_rate": 0.0001,
+      "loss": 3.0088,
+      "loss/crossentropy": 2.501011300086975,
+      "loss/hidden": 2.7625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2248419776558876,
+      "loss/reg": 0.0,
+      "step": 30900
+    },
+    {
+      "epoch": 0.20335526315789473,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.0226470947265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0529,
+      "loss/crossentropy": 2.4394679188728334,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23242839723825454,
+      "loss/reg": 0.0,
+      "step": 30910
+    },
+    {
+      "epoch": 0.20342105263157895,
+      "grad_norm": 1.9296875,
+      "grad_norm_var": 0.06285374959309896,
+      "learning_rate": 0.0001,
+      "loss": 2.9565,
+      "loss/crossentropy": 2.1788210391998293,
+      "loss/hidden": 2.690625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21422138065099716,
+      "loss/reg": 0.0,
+      "step": 30920
+    },
+    {
+      "epoch": 0.20348684210526316,
+      "grad_norm": 3.09375,
+      "grad_norm_var": 0.09987157185872396,
+      "learning_rate": 0.0001,
+      "loss": 2.948,
+      "loss/crossentropy": 2.4877074480056764,
+      "loss/hidden": 2.6375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22587471902370454,
+      "loss/reg": 0.0,
+      "step": 30930
+    },
+    {
+      "epoch": 0.20355263157894737,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.2575154622395833,
+      "learning_rate": 0.0001,
+      "loss": 3.0854,
+      "loss/crossentropy": 2.3099985003471373,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23227183222770692,
+      "loss/reg": 0.0,
+      "step": 30940
+    },
+    {
+      "epoch": 0.2036184210526316,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.18185221354166667,
+      "learning_rate": 0.0001,
+      "loss": 3.1055,
+      "loss/crossentropy": 2.228105306625366,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24457025676965713,
+      "loss/reg": 0.0,
+      "step": 30950
+    },
+    {
+      "epoch": 0.2036842105263158,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.08284098307291667,
+      "learning_rate": 0.0001,
+      "loss": 3.131,
+      "loss/crossentropy": 2.395846438407898,
+      "loss/hidden": 3.04375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2948496550321579,
+      "loss/reg": 0.0,
+      "step": 30960
+    },
+    {
+      "epoch": 0.20375,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.07164306640625,
+      "learning_rate": 0.0001,
+      "loss": 2.9739,
+      "loss/crossentropy": 2.221297824382782,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21414069309830666,
+      "loss/reg": 0.0,
+      "step": 30970
+    },
+    {
+      "epoch": 0.2038157894736842,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.13599853515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1104,
+      "loss/crossentropy": 2.293365275859833,
+      "loss/hidden": 2.9140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2863286077976227,
+      "loss/reg": 0.0,
+      "step": 30980
+    },
+    {
+      "epoch": 0.2038815789473684,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.12737630208333334,
+      "learning_rate": 0.0001,
+      "loss": 2.9774,
+      "loss/crossentropy": 2.293133610486984,
+      "loss/hidden": 2.7,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22788952738046647,
+      "loss/reg": 0.0,
+      "step": 30990
+    },
+    {
+      "epoch": 0.20394736842105263,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.01656494140625,
+      "learning_rate": 0.0001,
+      "loss": 2.9724,
+      "loss/crossentropy": 2.4067237973213196,
+      "loss/hidden": 2.584375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.220338836312294,
+      "loss/reg": 0.0,
+      "step": 31000
+    },
+    {
+      "epoch": 0.20401315789473684,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.984301503499349,
+      "learning_rate": 0.0001,
+      "loss": 3.0445,
+      "loss/crossentropy": 2.3435217142105103,
+      "loss/hidden": 2.9984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28915616124868393,
+      "loss/reg": 0.0,
+      "step": 31010
+    },
+    {
+      "epoch": 0.20407894736842105,
+      "grad_norm": 3.0,
+      "grad_norm_var": 0.10872573852539062,
+      "learning_rate": 0.0001,
+      "loss": 3.133,
+      "loss/crossentropy": 2.3546693086624146,
+      "loss/hidden": 2.8234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2506974846124649,
+      "loss/reg": 0.0,
+      "step": 31020
+    },
+    {
+      "epoch": 0.20414473684210527,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.33428929646809896,
+      "learning_rate": 0.0001,
+      "loss": 3.0211,
+      "loss/crossentropy": 2.2451566100120544,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23826195299625397,
+      "loss/reg": 0.0,
+      "step": 31030
+    },
+    {
+      "epoch": 0.20421052631578948,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.31133626302083334,
+      "learning_rate": 0.0001,
+      "loss": 3.0125,
+      "loss/crossentropy": 2.142116904258728,
+      "loss/hidden": 2.6671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19715375155210496,
+      "loss/reg": 0.0,
+      "step": 31040
+    },
+    {
+      "epoch": 0.2042763157894737,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.1289629618326823,
+      "learning_rate": 0.0001,
+      "loss": 3.0955,
+      "loss/crossentropy": 2.210948419570923,
+      "loss/hidden": 3.0921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2645356684923172,
+      "loss/reg": 0.0,
+      "step": 31050
+    },
+    {
+      "epoch": 0.2043421052631579,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.1048004150390625,
+      "learning_rate": 0.0001,
+      "loss": 3.1043,
+      "loss/crossentropy": 2.0977415561676027,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2552319660782814,
+      "loss/reg": 0.0,
+      "step": 31060
+    },
+    {
+      "epoch": 0.2044078947368421,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.08701070149739583,
+      "learning_rate": 0.0001,
+      "loss": 3.1067,
+      "loss/crossentropy": 2.264439880847931,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2655411049723625,
+      "loss/reg": 0.0,
+      "step": 31070
+    },
+    {
+      "epoch": 0.2044736842105263,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.06838785807291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0627,
+      "loss/crossentropy": 2.0332413136959078,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21882982477545737,
+      "loss/reg": 0.0,
+      "step": 31080
+    },
+    {
+      "epoch": 0.20453947368421052,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.10865478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0466,
+      "loss/crossentropy": 2.2589586079120636,
+      "loss/hidden": 2.6984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2315130352973938,
+      "loss/reg": 0.0,
+      "step": 31090
+    },
+    {
+      "epoch": 0.20460526315789473,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.37800191243489584,
+      "learning_rate": 0.0001,
+      "loss": 3.0036,
+      "loss/crossentropy": 2.3440281629562376,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2972772717475891,
+      "loss/reg": 0.0,
+      "step": 31100
+    },
+    {
+      "epoch": 0.20467105263157895,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.29983723958333336,
+      "learning_rate": 0.0001,
+      "loss": 2.954,
+      "loss/crossentropy": 2.47297340631485,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24424962252378463,
+      "loss/reg": 0.0,
+      "step": 31110
+    },
+    {
+      "epoch": 0.20473684210526316,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.2591634114583333,
+      "learning_rate": 0.0001,
+      "loss": 3.1546,
+      "loss/crossentropy": 2.436366784572601,
+      "loss/hidden": 2.9546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3093844935297966,
+      "loss/reg": 0.0,
+      "step": 31120
+    },
+    {
+      "epoch": 0.20480263157894738,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.2005859375,
+      "learning_rate": 0.0001,
+      "loss": 3.0745,
+      "loss/crossentropy": 2.140506219863892,
+      "loss/hidden": 2.8859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27134309709072113,
+      "loss/reg": 0.0,
+      "step": 31130
+    },
+    {
+      "epoch": 0.2048684210526316,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.032713826497395834,
+      "learning_rate": 0.0001,
+      "loss": 3.0449,
+      "loss/crossentropy": 2.5245323538780213,
+      "loss/hidden": 3.009375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2610665872693062,
+      "loss/reg": 0.0,
+      "step": 31140
+    },
+    {
+      "epoch": 0.2049342105263158,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.11193745930989583,
+      "learning_rate": 0.0001,
+      "loss": 3.0709,
+      "loss/crossentropy": 2.5337566137313843,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2527821347117424,
+      "loss/reg": 0.0,
+      "step": 31150
+    },
+    {
+      "epoch": 0.205,
+      "grad_norm": 3.953125,
+      "grad_norm_var": 0.3135943094889323,
+      "learning_rate": 0.0001,
+      "loss": 3.1177,
+      "loss/crossentropy": 2.2265261888504027,
+      "loss/hidden": 2.9515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2549376994371414,
+      "loss/reg": 0.0,
+      "step": 31160
+    },
+    {
+      "epoch": 0.2050657894736842,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.3199534098307292,
+      "learning_rate": 0.0001,
+      "loss": 3.0723,
+      "loss/crossentropy": 2.47224338054657,
+      "loss/hidden": 2.6328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2135899156332016,
+      "loss/reg": 0.0,
+      "step": 31170
+    },
+    {
+      "epoch": 0.20513157894736841,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.08787333170572917,
+      "learning_rate": 0.0001,
+      "loss": 3.006,
+      "loss/crossentropy": 1.9317480087280274,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1986733317375183,
+      "loss/reg": 0.0,
+      "step": 31180
+    },
+    {
+      "epoch": 0.20519736842105263,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.06467692057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0381,
+      "loss/crossentropy": 2.470042097568512,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2246125504374504,
+      "loss/reg": 0.0,
+      "step": 31190
+    },
+    {
+      "epoch": 0.20526315789473684,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.05276692708333333,
+      "learning_rate": 0.0001,
+      "loss": 2.9766,
+      "loss/crossentropy": 2.350657284259796,
+      "loss/hidden": 2.675,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22050067782402039,
+      "loss/reg": 0.0,
+      "step": 31200
+    },
+    {
+      "epoch": 0.20532894736842106,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.0873931884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0405,
+      "loss/crossentropy": 2.3005827188491823,
+      "loss/hidden": 2.734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2127441346645355,
+      "loss/reg": 0.0,
+      "step": 31210
+    },
+    {
+      "epoch": 0.20539473684210527,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 5.270894368489583,
+      "learning_rate": 0.0001,
+      "loss": 3.1111,
+      "loss/crossentropy": 2.4935895919799806,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2504632011055946,
+      "loss/reg": 0.0,
+      "step": 31220
+    },
+    {
+      "epoch": 0.20546052631578948,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.06770426432291667,
+      "learning_rate": 0.0001,
+      "loss": 2.9834,
+      "loss/crossentropy": 2.355222475528717,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2353107064962387,
+      "loss/reg": 0.0,
+      "step": 31230
+    },
+    {
+      "epoch": 0.2055263157894737,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.209130859375,
+      "learning_rate": 0.0001,
+      "loss": 3.0459,
+      "loss/crossentropy": 2.504118573665619,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22558754682540894,
+      "loss/reg": 0.0,
+      "step": 31240
+    },
+    {
+      "epoch": 0.20559210526315788,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.34091389973958336,
+      "learning_rate": 0.0001,
+      "loss": 3.0119,
+      "loss/crossentropy": 2.1399470806121825,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2328704759478569,
+      "loss/reg": 0.0,
+      "step": 31250
+    },
+    {
+      "epoch": 0.2056578947368421,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.24337565104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0416,
+      "loss/crossentropy": 2.251539409160614,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3552345484495163,
+      "loss/reg": 0.0,
+      "step": 31260
+    },
+    {
+      "epoch": 0.2057236842105263,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.28566080729166665,
+      "learning_rate": 0.0001,
+      "loss": 3.0568,
+      "loss/crossentropy": 2.310221529006958,
+      "loss/hidden": 2.940625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26540524512529373,
+      "loss/reg": 0.0,
+      "step": 31270
+    },
+    {
+      "epoch": 0.20578947368421052,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.18033447265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0158,
+      "loss/crossentropy": 2.489114725589752,
+      "loss/hidden": 2.96875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24016214907169342,
+      "loss/reg": 0.0,
+      "step": 31280
+    },
+    {
+      "epoch": 0.20585526315789474,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.09659830729166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0496,
+      "loss/crossentropy": 1.9956150293350219,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19284655302762985,
+      "loss/reg": 0.0,
+      "step": 31290
+    },
+    {
+      "epoch": 0.20592105263157895,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.0728179931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0451,
+      "loss/crossentropy": 2.383978569507599,
+      "loss/hidden": 2.7125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21748270690441132,
+      "loss/reg": 0.0,
+      "step": 31300
+    },
+    {
+      "epoch": 0.20598684210526316,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.05699462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.031,
+      "loss/crossentropy": 2.064331567287445,
+      "loss/hidden": 2.709375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21027912348508834,
+      "loss/reg": 0.0,
+      "step": 31310
+    },
+    {
+      "epoch": 0.20605263157894738,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.27751363118489586,
+      "learning_rate": 0.0001,
+      "loss": 3.0386,
+      "loss/crossentropy": 2.5503376483917237,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2552351266145706,
+      "loss/reg": 0.0,
+      "step": 31320
+    },
+    {
+      "epoch": 0.2061184210526316,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.15871480305989583,
+      "learning_rate": 0.0001,
+      "loss": 3.0885,
+      "loss/crossentropy": 2.408406615257263,
+      "loss/hidden": 2.71875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22667487859725952,
+      "loss/reg": 0.0,
+      "step": 31330
+    },
+    {
+      "epoch": 0.2061842105263158,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.11311009724934896,
+      "learning_rate": 0.0001,
+      "loss": 3.0571,
+      "loss/crossentropy": 2.199143981933594,
+      "loss/hidden": 2.821875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22705173045396804,
+      "loss/reg": 0.0,
+      "step": 31340
+    },
+    {
+      "epoch": 0.20625,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.11831868489583333,
+      "learning_rate": 0.0001,
+      "loss": 3.0539,
+      "loss/crossentropy": 1.9996395468711854,
+      "loss/hidden": 2.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22363494634628295,
+      "loss/reg": 0.0,
+      "step": 31350
+    },
+    {
+      "epoch": 0.2063157894736842,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.0868364969889323,
+      "learning_rate": 0.0001,
+      "loss": 3.053,
+      "loss/crossentropy": 2.067913568019867,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2136005848646164,
+      "loss/reg": 0.0,
+      "step": 31360
+    },
+    {
+      "epoch": 0.20638157894736842,
+      "grad_norm": 2.0,
+      "grad_norm_var": 0.086376953125,
+      "learning_rate": 0.0001,
+      "loss": 3.0064,
+      "loss/crossentropy": 2.300830841064453,
+      "loss/hidden": 2.809375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2269425220787525,
+      "loss/reg": 0.0,
+      "step": 31370
+    },
+    {
+      "epoch": 0.20644736842105263,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.05623753865559896,
+      "learning_rate": 0.0001,
+      "loss": 3.0327,
+      "loss/crossentropy": 2.283673417568207,
+      "loss/hidden": 2.775,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22951046377420425,
+      "loss/reg": 0.0,
+      "step": 31380
+    },
+    {
+      "epoch": 0.20651315789473684,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.15467020670572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0314,
+      "loss/crossentropy": 2.2721763372421266,
+      "loss/hidden": 2.70625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2015401691198349,
+      "loss/reg": 0.0,
+      "step": 31390
+    },
+    {
+      "epoch": 0.20657894736842106,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.17532958984375,
+      "learning_rate": 0.0001,
+      "loss": 3.0562,
+      "loss/crossentropy": 2.3926843643188476,
+      "loss/hidden": 2.6703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22229372709989548,
+      "loss/reg": 0.0,
+      "step": 31400
+    },
+    {
+      "epoch": 0.20664473684210527,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.16061986287434896,
+      "learning_rate": 0.0001,
+      "loss": 3.0372,
+      "loss/crossentropy": 2.2162522673606873,
+      "loss/hidden": 2.8,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21594754755496978,
+      "loss/reg": 0.0,
+      "step": 31410
+    },
+    {
+      "epoch": 0.20671052631578948,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.19730606079101562,
+      "learning_rate": 0.0001,
+      "loss": 2.9993,
+      "loss/crossentropy": 2.2487839818000794,
+      "loss/hidden": 2.6734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23367298394441605,
+      "loss/reg": 0.0,
+      "step": 31420
+    },
+    {
+      "epoch": 0.2067763157894737,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.22303059895833333,
+      "learning_rate": 0.0001,
+      "loss": 3.03,
+      "loss/crossentropy": 2.304809939861298,
+      "loss/hidden": 2.6609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21387835443019867,
+      "loss/reg": 0.0,
+      "step": 31430
+    },
+    {
+      "epoch": 0.20684210526315788,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.18980204264322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0591,
+      "loss/crossentropy": 2.428087902069092,
+      "loss/hidden": 2.65625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2132945567369461,
+      "loss/reg": 0.0,
+      "step": 31440
+    },
+    {
+      "epoch": 0.2069078947368421,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.11033426920572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0247,
+      "loss/crossentropy": 2.3415855526924134,
+      "loss/hidden": 2.7296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2590311750769615,
+      "loss/reg": 0.0,
+      "step": 31450
+    },
+    {
+      "epoch": 0.2069736842105263,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 2.33717041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0733,
+      "loss/crossentropy": 2.0509816646575927,
+      "loss/hidden": 2.7125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22177468091249466,
+      "loss/reg": 0.0,
+      "step": 31460
+    },
+    {
+      "epoch": 0.20703947368421052,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.06643473307291667,
+      "learning_rate": 0.0001,
+      "loss": 3.01,
+      "loss/crossentropy": 2.0959688067436217,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2369927003979683,
+      "loss/reg": 0.0,
+      "step": 31470
+    },
+    {
+      "epoch": 0.20710526315789474,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.07058690388997396,
+      "learning_rate": 0.0001,
+      "loss": 3.0335,
+      "loss/crossentropy": 2.2887695908546446,
+      "loss/hidden": 2.7125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21632779389619827,
+      "loss/reg": 0.0,
+      "step": 31480
+    },
+    {
+      "epoch": 0.20717105263157895,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.0905914306640625,
+      "learning_rate": 0.0001,
+      "loss": 3.022,
+      "loss/crossentropy": 2.251658821105957,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27912095934152603,
+      "loss/reg": 0.0,
+      "step": 31490
+    },
+    {
+      "epoch": 0.20723684210526316,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.13886617024739584,
+      "learning_rate": 0.0001,
+      "loss": 3.0013,
+      "loss/crossentropy": 2.200524830818176,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20394676253199578,
+      "loss/reg": 0.0,
+      "step": 31500
+    },
+    {
+      "epoch": 0.20730263157894738,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.11672770182291667,
+      "learning_rate": 0.0001,
+      "loss": 3.1006,
+      "loss/crossentropy": 2.410679817199707,
+      "loss/hidden": 2.7390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2168733671307564,
+      "loss/reg": 0.0,
+      "step": 31510
+    },
+    {
+      "epoch": 0.2073684210526316,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.16353759765625,
+      "learning_rate": 0.0001,
+      "loss": 2.9942,
+      "loss/crossentropy": 2.1653492391109466,
+      "loss/hidden": 2.5796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18678945749998094,
+      "loss/reg": 0.0,
+      "step": 31520
+    },
+    {
+      "epoch": 0.20743421052631578,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.22099202473958332,
+      "learning_rate": 0.0001,
+      "loss": 3.0174,
+      "loss/crossentropy": 2.4728673577308653,
+      "loss/hidden": 2.7515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24382270276546478,
+      "loss/reg": 0.0,
+      "step": 31530
+    },
+    {
+      "epoch": 0.2075,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.20640869140625,
+      "learning_rate": 0.0001,
+      "loss": 3.065,
+      "loss/crossentropy": 2.3791961908340453,
+      "loss/hidden": 3.0109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29753602892160413,
+      "loss/reg": 0.0,
+      "step": 31540
+    },
+    {
+      "epoch": 0.2075657894736842,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.07801005045572916,
+      "learning_rate": 0.0001,
+      "loss": 2.9835,
+      "loss/crossentropy": 2.1349298536777495,
+      "loss/hidden": 2.6390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2179326094686985,
+      "loss/reg": 0.0,
+      "step": 31550
+    },
+    {
+      "epoch": 0.20763157894736842,
+      "grad_norm": 1.9921875,
+      "grad_norm_var": 0.3508989969889323,
+      "learning_rate": 0.0001,
+      "loss": 3.0481,
+      "loss/crossentropy": 2.128028416633606,
+      "loss/hidden": 2.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24575520306825638,
+      "loss/reg": 0.0,
+      "step": 31560
+    },
+    {
+      "epoch": 0.20769736842105263,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.36219253540039065,
+      "learning_rate": 0.0001,
+      "loss": 2.992,
+      "loss/crossentropy": 2.366794526576996,
+      "loss/hidden": 2.7875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2530692145228386,
+      "loss/reg": 0.0,
+      "step": 31570
+    },
+    {
+      "epoch": 0.20776315789473684,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.18565165201822917,
+      "learning_rate": 0.0001,
+      "loss": 2.9888,
+      "loss/crossentropy": 2.5053478240966798,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24881082475185395,
+      "loss/reg": 0.0,
+      "step": 31580
+    },
+    {
+      "epoch": 0.20782894736842106,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.22190348307291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0479,
+      "loss/crossentropy": 2.115998589992523,
+      "loss/hidden": 2.8984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2497893139719963,
+      "loss/reg": 0.0,
+      "step": 31590
+    },
+    {
+      "epoch": 0.20789473684210527,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.0385406494140625,
+      "learning_rate": 0.0001,
+      "loss": 2.9967,
+      "loss/crossentropy": 2.2767433404922484,
+      "loss/hidden": 2.6578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22527251839637757,
+      "loss/reg": 0.0,
+      "step": 31600
+    },
+    {
+      "epoch": 0.20796052631578948,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.11295166015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0751,
+      "loss/crossentropy": 2.271492937207222,
+      "loss/hidden": 2.78125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2081324838101864,
+      "loss/reg": 0.0,
+      "step": 31610
+    },
+    {
+      "epoch": 0.2080263157894737,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.1297271728515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0283,
+      "loss/crossentropy": 2.107327163219452,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23592574074864386,
+      "loss/reg": 0.0,
+      "step": 31620
+    },
+    {
+      "epoch": 0.20809210526315788,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.06132583618164063,
+      "learning_rate": 0.0001,
+      "loss": 3.0553,
+      "loss/crossentropy": 2.4657732486724853,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2265238046646118,
+      "loss/reg": 0.0,
+      "step": 31630
+    },
+    {
+      "epoch": 0.2081578947368421,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.5120076497395833,
+      "learning_rate": 0.0001,
+      "loss": 3.0358,
+      "loss/crossentropy": 2.4768943071365355,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2473340943455696,
+      "loss/reg": 0.0,
+      "step": 31640
+    },
+    {
+      "epoch": 0.2082236842105263,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.35252176920572914,
+      "learning_rate": 0.0001,
+      "loss": 3.0949,
+      "loss/crossentropy": 2.5071337461471557,
+      "loss/hidden": 2.678125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23168605715036392,
+      "loss/reg": 0.0,
+      "step": 31650
+    },
+    {
+      "epoch": 0.20828947368421052,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.19873758951822917,
+      "learning_rate": 0.0001,
+      "loss": 3.0033,
+      "loss/crossentropy": 2.3884094834327696,
+      "loss/hidden": 2.7515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2127884179353714,
+      "loss/reg": 0.0,
+      "step": 31660
+    },
+    {
+      "epoch": 0.20835526315789474,
+      "grad_norm": 3.421875,
+      "grad_norm_var": 0.33858133951822916,
+      "learning_rate": 0.0001,
+      "loss": 2.9746,
+      "loss/crossentropy": 2.172113299369812,
+      "loss/hidden": 2.6140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20361884087324142,
+      "loss/reg": 0.0,
+      "step": 31670
+    },
+    {
+      "epoch": 0.20842105263157895,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.36571858723958334,
+      "learning_rate": 0.0001,
+      "loss": 2.9778,
+      "loss/crossentropy": 2.3501265048980713,
+      "loss/hidden": 2.696875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21666382998228073,
+      "loss/reg": 0.0,
+      "step": 31680
+    },
+    {
+      "epoch": 0.20848684210526316,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.12886962890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1071,
+      "loss/crossentropy": 2.8127560138702394,
+      "loss/hidden": 3.221875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25830034017562864,
+      "loss/reg": 0.0,
+      "step": 31690
+    },
+    {
+      "epoch": 0.20855263157894738,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.04257405598958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0355,
+      "loss/crossentropy": 2.0618197679519654,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22268864512443542,
+      "loss/reg": 0.0,
+      "step": 31700
+    },
+    {
+      "epoch": 0.2086184210526316,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.10972391764322917,
+      "learning_rate": 0.0001,
+      "loss": 2.9789,
+      "loss/crossentropy": 2.4682215332984923,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27270212322473525,
+      "loss/reg": 0.0,
+      "step": 31710
+    },
+    {
+      "epoch": 0.20868421052631578,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.108251953125,
+      "learning_rate": 0.0001,
+      "loss": 3.0187,
+      "loss/crossentropy": 2.3017919063568115,
+      "loss/hidden": 2.6421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2289327010512352,
+      "loss/reg": 0.0,
+      "step": 31720
+    },
+    {
+      "epoch": 0.20875,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.08994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0376,
+      "loss/crossentropy": 2.2758594751358032,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23465103954076766,
+      "loss/reg": 0.0,
+      "step": 31730
+    },
+    {
+      "epoch": 0.2088157894736842,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.0693511962890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9794,
+      "loss/crossentropy": 2.3359742760658264,
+      "loss/hidden": 2.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25490144789218905,
+      "loss/reg": 0.0,
+      "step": 31740
+    },
+    {
+      "epoch": 0.20888157894736842,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.035399373372395834,
+      "learning_rate": 0.0001,
+      "loss": 3.0153,
+      "loss/crossentropy": 2.2797624588012697,
+      "loss/hidden": 2.6109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1947791814804077,
+      "loss/reg": 0.0,
+      "step": 31750
+    },
+    {
+      "epoch": 0.20894736842105263,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.048460896809895834,
+      "learning_rate": 0.0001,
+      "loss": 3.0293,
+      "loss/crossentropy": 2.221743679046631,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2385573446750641,
+      "loss/reg": 0.0,
+      "step": 31760
+    },
+    {
+      "epoch": 0.20901315789473685,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.0992828369140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0352,
+      "loss/crossentropy": 2.3273940563201903,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22714275866746902,
+      "loss/reg": 0.0,
+      "step": 31770
+    },
+    {
+      "epoch": 0.20907894736842106,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.05660400390625,
+      "learning_rate": 0.0001,
+      "loss": 2.9783,
+      "loss/crossentropy": 2.5335972189903258,
+      "loss/hidden": 2.653125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22051512748003005,
+      "loss/reg": 0.0,
+      "step": 31780
+    },
+    {
+      "epoch": 0.20914473684210527,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.0770172119140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0626,
+      "loss/crossentropy": 2.1290286660194395,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24563827514648437,
+      "loss/reg": 0.0,
+      "step": 31790
+    },
+    {
+      "epoch": 0.20921052631578949,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.12021484375,
+      "learning_rate": 0.0001,
+      "loss": 3.1026,
+      "loss/crossentropy": 2.277496612071991,
+      "loss/hidden": 2.8890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23920933827757834,
+      "loss/reg": 0.0,
+      "step": 31800
+    },
+    {
+      "epoch": 0.20927631578947367,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.07742513020833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0043,
+      "loss/crossentropy": 2.3269221425056457,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2029653638601303,
+      "loss/reg": 0.0,
+      "step": 31810
+    },
+    {
+      "epoch": 0.20934210526315788,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.02303034464518229,
+      "learning_rate": 0.0001,
+      "loss": 3.0615,
+      "loss/crossentropy": 2.352055883407593,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2182576224207878,
+      "loss/reg": 0.0,
+      "step": 31820
+    },
+    {
+      "epoch": 0.2094078947368421,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.048500315348307295,
+      "learning_rate": 0.0001,
+      "loss": 3.0806,
+      "loss/crossentropy": 2.143119287490845,
+      "loss/hidden": 2.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24638065993785857,
+      "loss/reg": 0.0,
+      "step": 31830
+    },
+    {
+      "epoch": 0.2094736842105263,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.13288472493489584,
+      "learning_rate": 0.0001,
+      "loss": 3.1094,
+      "loss/crossentropy": 2.344961977005005,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24402716457843782,
+      "loss/reg": 0.0,
+      "step": 31840
+    },
+    {
+      "epoch": 0.20953947368421053,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.39104817708333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0091,
+      "loss/crossentropy": 2.3289579272270204,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21923962533473967,
+      "loss/reg": 0.0,
+      "step": 31850
+    },
+    {
+      "epoch": 0.20960526315789474,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.35844624837239586,
+      "learning_rate": 0.0001,
+      "loss": 3.1472,
+      "loss/crossentropy": 2.2578481793403626,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2393498733639717,
+      "loss/reg": 0.0,
+      "step": 31860
+    },
+    {
+      "epoch": 0.20967105263157895,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.103369140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0043,
+      "loss/crossentropy": 2.1479632019996644,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21754586473107337,
+      "loss/reg": 0.0,
+      "step": 31870
+    },
+    {
+      "epoch": 0.20973684210526317,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.05641988118489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0913,
+      "loss/crossentropy": 2.31242595911026,
+      "loss/hidden": 3.0078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3139140591025352,
+      "loss/reg": 0.0,
+      "step": 31880
+    },
+    {
+      "epoch": 0.20980263157894738,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.14700113932291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0276,
+      "loss/crossentropy": 2.421666181087494,
+      "loss/hidden": 2.703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21645313799381255,
+      "loss/reg": 0.0,
+      "step": 31890
+    },
+    {
+      "epoch": 0.20986842105263157,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.6790028889973958,
+      "learning_rate": 0.0001,
+      "loss": 3.0781,
+      "loss/crossentropy": 2.2751805901527407,
+      "loss/hidden": 2.89375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22757814675569535,
+      "loss/reg": 0.0,
+      "step": 31900
+    },
+    {
+      "epoch": 0.20993421052631578,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.14388020833333334,
+      "learning_rate": 0.0001,
+      "loss": 2.978,
+      "loss/crossentropy": 2.3446357250213623,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21789460927248,
+      "loss/reg": 0.0,
+      "step": 31910
+    },
+    {
+      "epoch": 0.21,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.1076324462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0473,
+      "loss/crossentropy": 2.3983862519264223,
+      "loss/hidden": 2.6640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21307818740606307,
+      "loss/reg": 0.0,
+      "step": 31920
+    },
+    {
+      "epoch": 0.2100657894736842,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.03589579264322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0171,
+      "loss/crossentropy": 2.37126282453537,
+      "loss/hidden": 2.7,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.205704665184021,
+      "loss/reg": 0.0,
+      "step": 31930
+    },
+    {
+      "epoch": 0.21013157894736842,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.08915608723958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0071,
+      "loss/crossentropy": 2.2473104357719422,
+      "loss/hidden": 2.6234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2207282856106758,
+      "loss/reg": 0.0,
+      "step": 31940
+    },
+    {
+      "epoch": 0.21019736842105263,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.03990478515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0541,
+      "loss/crossentropy": 2.4488544821739198,
+      "loss/hidden": 2.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26014447808265684,
+      "loss/reg": 0.0,
+      "step": 31950
+    },
+    {
+      "epoch": 0.21026315789473685,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.0229156494140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0762,
+      "loss/crossentropy": 1.9240918695926665,
+      "loss/hidden": 2.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19550290815532206,
+      "loss/reg": 0.0,
+      "step": 31960
+    },
+    {
+      "epoch": 0.21032894736842106,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.03746337890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0402,
+      "loss/crossentropy": 2.1934541881084444,
+      "loss/hidden": 2.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19934962689876556,
+      "loss/reg": 0.0,
+      "step": 31970
+    },
+    {
+      "epoch": 0.21039473684210527,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.03241780598958333,
+      "learning_rate": 0.0001,
+      "loss": 2.9459,
+      "loss/crossentropy": 2.071032130718231,
+      "loss/hidden": 2.61875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20514383763074875,
+      "loss/reg": 0.0,
+      "step": 31980
+    },
+    {
+      "epoch": 0.2104605263157895,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.02534764607747396,
+      "learning_rate": 0.0001,
+      "loss": 3.0513,
+      "loss/crossentropy": 2.274570310115814,
+      "loss/hidden": 2.9515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3155985027551651,
+      "loss/reg": 0.0,
+      "step": 31990
+    },
+    {
+      "epoch": 0.21052631578947367,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 3.301877391722643e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1093,
+      "loss/crossentropy": 2.0131736040115356,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2141942039132118,
+      "loss/reg": 0.0,
+      "step": 32000
+    },
+    {
+      "epoch": 0.21059210526315789,
+      "grad_norm": 3.078125,
+      "grad_norm_var": 0.2028717041015625,
+      "learning_rate": 0.0001,
+      "loss": 2.9957,
+      "loss/crossentropy": 2.213227319717407,
+      "loss/hidden": 2.6984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22059416845440866,
+      "loss/reg": 0.0,
+      "step": 32010
+    },
+    {
+      "epoch": 0.2106578947368421,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.09852676391601563,
+      "learning_rate": 0.0001,
+      "loss": 2.9318,
+      "loss/crossentropy": 2.2327592849731444,
+      "loss/hidden": 2.6125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19538245126605033,
+      "loss/reg": 0.0,
+      "step": 32020
+    },
+    {
+      "epoch": 0.2107236842105263,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.10897191365559895,
+      "learning_rate": 0.0001,
+      "loss": 2.9561,
+      "loss/crossentropy": 2.415834832191467,
+      "loss/hidden": 2.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2341999500989914,
+      "loss/reg": 0.0,
+      "step": 32030
+    },
+    {
+      "epoch": 0.21078947368421053,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.0959185282389323,
+      "learning_rate": 0.0001,
+      "loss": 2.9979,
+      "loss/crossentropy": 2.4787596464157104,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23873308449983596,
+      "loss/reg": 0.0,
+      "step": 32040
+    },
+    {
+      "epoch": 0.21085526315789474,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.12280985514322916,
+      "learning_rate": 0.0001,
+      "loss": 2.9578,
+      "loss/crossentropy": 2.3683923482894897,
+      "loss/hidden": 2.653125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22164048850536347,
+      "loss/reg": 0.0,
+      "step": 32050
+    },
+    {
+      "epoch": 0.21092105263157895,
+      "grad_norm": 4.40625,
+      "grad_norm_var": 0.3617828369140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0122,
+      "loss/crossentropy": 2.336903250217438,
+      "loss/hidden": 2.671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2218891516327858,
+      "loss/reg": 0.0,
+      "step": 32060
+    },
+    {
+      "epoch": 0.21098684210526317,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.31454671223958336,
+      "learning_rate": 0.0001,
+      "loss": 2.9132,
+      "loss/crossentropy": 2.3459483861923216,
+      "loss/hidden": 2.715625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20081116408109664,
+      "loss/reg": 0.0,
+      "step": 32070
+    },
+    {
+      "epoch": 0.21105263157894738,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.04011942545572917,
+      "learning_rate": 0.0001,
+      "loss": 2.9616,
+      "loss/crossentropy": 2.1744574666023255,
+      "loss/hidden": 2.7125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21101347357034683,
+      "loss/reg": 0.0,
+      "step": 32080
+    },
+    {
+      "epoch": 0.21111842105263157,
+      "grad_norm": 3.921875,
+      "grad_norm_var": 0.22661107381184895,
+      "learning_rate": 0.0001,
+      "loss": 2.9912,
+      "loss/crossentropy": 2.1881649017333986,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2135306864976883,
+      "loss/reg": 0.0,
+      "step": 32090
+    },
+    {
+      "epoch": 0.21118421052631578,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.20420506795247395,
+      "learning_rate": 0.0001,
+      "loss": 3.0331,
+      "loss/crossentropy": 2.0982267916202546,
+      "loss/hidden": 2.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20456733107566832,
+      "loss/reg": 0.0,
+      "step": 32100
+    },
+    {
+      "epoch": 0.21125,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.12456766764322917,
+      "learning_rate": 0.0001,
+      "loss": 2.9625,
+      "loss/crossentropy": 2.291160595417023,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2443986713886261,
+      "loss/reg": 0.0,
+      "step": 32110
+    },
+    {
+      "epoch": 0.2113157894736842,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.3302968343098958,
+      "learning_rate": 0.0001,
+      "loss": 3.0737,
+      "loss/crossentropy": 2.3988569140434266,
+      "loss/hidden": 2.7109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22473453879356384,
+      "loss/reg": 0.0,
+      "step": 32120
+    },
+    {
+      "epoch": 0.21138157894736842,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 1.0375,
+      "learning_rate": 0.0001,
+      "loss": 2.9681,
+      "loss/crossentropy": 2.157117784023285,
+      "loss/hidden": 2.9140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25956482589244845,
+      "loss/reg": 0.0,
+      "step": 32130
+    },
+    {
+      "epoch": 0.21144736842105263,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.967632802327474,
+      "learning_rate": 0.0001,
+      "loss": 3.0129,
+      "loss/crossentropy": 2.007625603675842,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21298689395189285,
+      "loss/reg": 0.0,
+      "step": 32140
+    },
+    {
+      "epoch": 0.21151315789473685,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.09070536295572916,
+      "learning_rate": 0.0001,
+      "loss": 2.9862,
+      "loss/crossentropy": 2.002223217487335,
+      "loss/hidden": 2.584375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18768413811922074,
+      "loss/reg": 0.0,
+      "step": 32150
+    },
+    {
+      "epoch": 0.21157894736842106,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.04879557291666667,
+      "learning_rate": 0.0001,
+      "loss": 3.0679,
+      "loss/crossentropy": 2.470254373550415,
+      "loss/hidden": 2.8578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2670274436473846,
+      "loss/reg": 0.0,
+      "step": 32160
+    },
+    {
+      "epoch": 0.21164473684210527,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.04566650390625,
+      "learning_rate": 0.0001,
+      "loss": 2.9691,
+      "loss/crossentropy": 2.489493703842163,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2385897770524025,
+      "loss/reg": 0.0,
+      "step": 32170
+    },
+    {
+      "epoch": 0.21171052631578946,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.051512654622395834,
+      "learning_rate": 0.0001,
+      "loss": 3.079,
+      "loss/crossentropy": 2.4059490084648134,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2276516616344452,
+      "loss/reg": 0.0,
+      "step": 32180
+    },
+    {
+      "epoch": 0.21177631578947367,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.18515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9959,
+      "loss/crossentropy": 2.1902307987213137,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22064511626958846,
+      "loss/reg": 0.0,
+      "step": 32190
+    },
+    {
+      "epoch": 0.2118421052631579,
+      "grad_norm": 3.09375,
+      "grad_norm_var": 0.21318359375,
+      "learning_rate": 0.0001,
+      "loss": 3.0886,
+      "loss/crossentropy": 2.2975401639938355,
+      "loss/hidden": 2.790625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21864379346370696,
+      "loss/reg": 0.0,
+      "step": 32200
+    },
+    {
+      "epoch": 0.2119078947368421,
+      "grad_norm": 3.65625,
+      "grad_norm_var": 0.3596425374348958,
+      "learning_rate": 0.0001,
+      "loss": 3.0936,
+      "loss/crossentropy": 2.3152296900749207,
+      "loss/hidden": 2.64375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21215650141239167,
+      "loss/reg": 0.0,
+      "step": 32210
+    },
+    {
+      "epoch": 0.21197368421052631,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.4088053385416667,
+      "learning_rate": 0.0001,
+      "loss": 3.0212,
+      "loss/crossentropy": 2.345796763896942,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22906456291675567,
+      "loss/reg": 0.0,
+      "step": 32220
+    },
+    {
+      "epoch": 0.21203947368421053,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.36056315104166664,
+      "learning_rate": 0.0001,
+      "loss": 2.9913,
+      "loss/crossentropy": 2.2359737753868103,
+      "loss/hidden": 2.8265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21989362984895705,
+      "loss/reg": 0.0,
+      "step": 32230
+    },
+    {
+      "epoch": 0.21210526315789474,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.1499908447265625,
+      "learning_rate": 0.0001,
+      "loss": 2.9931,
+      "loss/crossentropy": 2.3540435433387756,
+      "loss/hidden": 2.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20313581377267836,
+      "loss/reg": 0.0,
+      "step": 32240
+    },
+    {
+      "epoch": 0.21217105263157895,
+      "grad_norm": 3.171875,
+      "grad_norm_var": 0.08063151041666666,
+      "learning_rate": 0.0001,
+      "loss": 2.966,
+      "loss/crossentropy": 2.3999782681465147,
+      "loss/hidden": 2.6953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20714087784290314,
+      "loss/reg": 0.0,
+      "step": 32250
+    },
+    {
+      "epoch": 0.21223684210526317,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.08381245930989584,
+      "learning_rate": 0.0001,
+      "loss": 3.0748,
+      "loss/crossentropy": 2.40082049369812,
+      "loss/hidden": 2.84375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24139860570430755,
+      "loss/reg": 0.0,
+      "step": 32260
+    },
+    {
+      "epoch": 0.21230263157894738,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.0259185791015625,
+      "learning_rate": 0.0001,
+      "loss": 2.9642,
+      "loss/crossentropy": 2.5221059322357178,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24260111004114152,
+      "loss/reg": 0.0,
+      "step": 32270
+    },
+    {
+      "epoch": 0.21236842105263157,
+      "grad_norm": 3.421875,
+      "grad_norm_var": 5.511455956857455e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1235,
+      "loss/crossentropy": 2.582364892959595,
+      "loss/hidden": 2.6109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2309617057442665,
+      "loss/reg": 0.0,
+      "step": 32280
+    },
+    {
+      "epoch": 0.21243421052631578,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 8.515180979484448e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1265,
+      "loss/crossentropy": 1.9406600832939147,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25122445076704025,
+      "loss/reg": 0.0,
+      "step": 32290
+    },
+    {
+      "epoch": 0.2125,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.11577860514322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0392,
+      "loss/crossentropy": 1.9736381709575652,
+      "loss/hidden": 2.94375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24801484644412994,
+      "loss/reg": 0.0,
+      "step": 32300
+    },
+    {
+      "epoch": 0.2125657894736842,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.11686909993489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0,
+      "loss/crossentropy": 2.57318320274353,
+      "loss/hidden": 2.7625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25067331790924074,
+      "loss/reg": 0.0,
+      "step": 32310
+    },
+    {
+      "epoch": 0.21263157894736842,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.11437352498372395,
+      "learning_rate": 0.0001,
+      "loss": 2.9663,
+      "loss/crossentropy": 2.3444815397262575,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23782878071069719,
+      "loss/reg": 0.0,
+      "step": 32320
+    },
+    {
+      "epoch": 0.21269736842105263,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.05796483357747396,
+      "learning_rate": 0.0001,
+      "loss": 2.9547,
+      "loss/crossentropy": 2.2045453786849976,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2417543426156044,
+      "loss/reg": 0.0,
+      "step": 32330
+    },
+    {
+      "epoch": 0.21276315789473685,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.045344034830729164,
+      "learning_rate": 0.0001,
+      "loss": 3.0757,
+      "loss/crossentropy": 2.349012243747711,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2658751055598259,
+      "loss/reg": 0.0,
+      "step": 32340
+    },
+    {
+      "epoch": 0.21282894736842106,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.04952799479166667,
+      "learning_rate": 0.0001,
+      "loss": 2.9861,
+      "loss/crossentropy": 2.4626625537872315,
+      "loss/hidden": 2.8953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25371641367673875,
+      "loss/reg": 0.0,
+      "step": 32350
+    },
+    {
+      "epoch": 0.21289473684210528,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.06630223592122396,
+      "learning_rate": 0.0001,
+      "loss": 2.9574,
+      "loss/crossentropy": 2.186726263165474,
+      "loss/hidden": 2.7625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20951557606458665,
+      "loss/reg": 0.0,
+      "step": 32360
+    },
+    {
+      "epoch": 0.21296052631578946,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.10197728474934896,
+      "learning_rate": 0.0001,
+      "loss": 2.965,
+      "loss/crossentropy": 2.340285396575928,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23474745452404022,
+      "loss/reg": 0.0,
+      "step": 32370
+    },
+    {
+      "epoch": 0.21302631578947367,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.157373046875,
+      "learning_rate": 0.0001,
+      "loss": 3.0962,
+      "loss/crossentropy": 2.47646986246109,
+      "loss/hidden": 2.9375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26605418920516966,
+      "loss/reg": 0.0,
+      "step": 32380
+    },
+    {
+      "epoch": 0.2130921052631579,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.10019429524739583,
+      "learning_rate": 0.0001,
+      "loss": 2.9578,
+      "loss/crossentropy": 1.7385891020298003,
+      "loss/hidden": 2.66875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.17920974977314472,
+      "loss/reg": 0.0,
+      "step": 32390
+    },
+    {
+      "epoch": 0.2131578947368421,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.14401041666666667,
+      "learning_rate": 0.0001,
+      "loss": 2.9707,
+      "loss/crossentropy": 2.182602137327194,
+      "loss/hidden": 2.7140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2166632428765297,
+      "loss/reg": 0.0,
+      "step": 32400
+    },
+    {
+      "epoch": 0.21322368421052632,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.15817057291666667,
+      "learning_rate": 0.0001,
+      "loss": 3.0277,
+      "loss/crossentropy": 2.3320940494537354,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.263668055832386,
+      "loss/reg": 0.0,
+      "step": 32410
+    },
+    {
+      "epoch": 0.21328947368421053,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.156201171875,
+      "learning_rate": 0.0001,
+      "loss": 3.0197,
+      "loss/crossentropy": 2.338312292098999,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22362788915634155,
+      "loss/reg": 0.0,
+      "step": 32420
+    },
+    {
+      "epoch": 0.21335526315789474,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.376025390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0193,
+      "loss/crossentropy": 2.3314969301223756,
+      "loss/hidden": 2.7078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2156234934926033,
+      "loss/reg": 0.0,
+      "step": 32430
+    },
+    {
+      "epoch": 0.21342105263157896,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.6317545572916666,
+      "learning_rate": 0.0001,
+      "loss": 2.97,
+      "loss/crossentropy": 2.3391250133514405,
+      "loss/hidden": 2.734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21457959488034248,
+      "loss/reg": 0.0,
+      "step": 32440
+    },
+    {
+      "epoch": 0.21348684210526317,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.3609375,
+      "learning_rate": 0.0001,
+      "loss": 3.0389,
+      "loss/crossentropy": 2.267109489440918,
+      "loss/hidden": 3.1234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3404072761535645,
+      "loss/reg": 0.0,
+      "step": 32450
+    },
+    {
+      "epoch": 0.21355263157894736,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.06378580729166666,
+      "learning_rate": 0.0001,
+      "loss": 2.9758,
+      "loss/crossentropy": 2.3441552996635435,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24242025315761567,
+      "loss/reg": 0.0,
+      "step": 32460
+    },
+    {
+      "epoch": 0.21361842105263157,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.0823883056640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0308,
+      "loss/crossentropy": 2.2937331914901735,
+      "loss/hidden": 2.7515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22603273689746856,
+      "loss/reg": 0.0,
+      "step": 32470
+    },
+    {
+      "epoch": 0.21368421052631578,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.14341532389322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0095,
+      "loss/crossentropy": 2.071279937028885,
+      "loss/hidden": 2.7390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1986673153936863,
+      "loss/reg": 0.0,
+      "step": 32480
+    },
+    {
+      "epoch": 0.21375,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.11785481770833334,
+      "learning_rate": 0.0001,
+      "loss": 3.0166,
+      "loss/crossentropy": 2.0449776887893676,
+      "loss/hidden": 2.9375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2660417690873146,
+      "loss/reg": 0.0,
+      "step": 32490
+    },
+    {
+      "epoch": 0.2138157894736842,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.07841389973958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0209,
+      "loss/crossentropy": 2.22779695391655,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2454911906272173,
+      "loss/reg": 0.0,
+      "step": 32500
+    },
+    {
+      "epoch": 0.21388157894736842,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.22735087076822916,
+      "learning_rate": 0.0001,
+      "loss": 3.0449,
+      "loss/crossentropy": 2.1109942197799683,
+      "loss/hidden": 3.0140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25361132323741914,
+      "loss/reg": 0.0,
+      "step": 32510
+    },
+    {
+      "epoch": 0.21394736842105264,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.21687825520833334,
+      "learning_rate": 0.0001,
+      "loss": 2.9946,
+      "loss/crossentropy": 2.594680833816528,
+      "loss/hidden": 2.734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24347000420093537,
+      "loss/reg": 0.0,
+      "step": 32520
+    },
+    {
+      "epoch": 0.21401315789473685,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.027469889322916666,
+      "learning_rate": 0.0001,
+      "loss": 2.935,
+      "loss/crossentropy": 2.4400468468666077,
+      "loss/hidden": 2.734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2244595929980278,
+      "loss/reg": 0.0,
+      "step": 32530
+    },
+    {
+      "epoch": 0.21407894736842106,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.02568359375,
+      "learning_rate": 0.0001,
+      "loss": 2.9661,
+      "loss/crossentropy": 2.336955714225769,
+      "loss/hidden": 2.6265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1992204010486603,
+      "loss/reg": 0.0,
+      "step": 32540
+    },
+    {
+      "epoch": 0.21414473684210528,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.051488240559895836,
+      "learning_rate": 0.0001,
+      "loss": 3.0259,
+      "loss/crossentropy": 2.2532342314720153,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26036129891872406,
+      "loss/reg": 0.0,
+      "step": 32550
+    },
+    {
+      "epoch": 0.21421052631578946,
+      "grad_norm": 1.9765625,
+      "grad_norm_var": 0.03230158487955729,
+      "learning_rate": 0.0001,
+      "loss": 2.9443,
+      "loss/crossentropy": 2.112717604637146,
+      "loss/hidden": 2.5703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.17500633224844933,
+      "loss/reg": 0.0,
+      "step": 32560
+    },
+    {
+      "epoch": 0.21427631578947368,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.03321711222330729,
+      "learning_rate": 0.0001,
+      "loss": 3.0203,
+      "loss/crossentropy": 2.288592982292175,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22093368023633958,
+      "loss/reg": 0.0,
+      "step": 32570
+    },
+    {
+      "epoch": 0.2143421052631579,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.32156575520833336,
+      "learning_rate": 0.0001,
+      "loss": 2.9199,
+      "loss/crossentropy": 2.1705434560775756,
+      "loss/hidden": 2.709375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20051559060811996,
+      "loss/reg": 0.0,
+      "step": 32580
+    },
+    {
+      "epoch": 0.2144078947368421,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.06936442057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0023,
+      "loss/crossentropy": 2.0764291286468506,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20810510888695716,
+      "loss/reg": 0.0,
+      "step": 32590
+    },
+    {
+      "epoch": 0.21447368421052632,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.09537760416666667,
+      "learning_rate": 0.0001,
+      "loss": 2.9267,
+      "loss/crossentropy": 2.377918064594269,
+      "loss/hidden": 2.8390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23365774154663085,
+      "loss/reg": 0.0,
+      "step": 32600
+    },
+    {
+      "epoch": 0.21453947368421053,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.13495686848958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0409,
+      "loss/crossentropy": 2.418037164211273,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2420778289437294,
+      "loss/reg": 0.0,
+      "step": 32610
+    },
+    {
+      "epoch": 0.21460526315789474,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.37043863932291665,
+      "learning_rate": 0.0001,
+      "loss": 2.9946,
+      "loss/crossentropy": 2.272017073631287,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.213658407330513,
+      "loss/reg": 0.0,
+      "step": 32620
+    },
+    {
+      "epoch": 0.21467105263157896,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.40995686848958335,
+      "learning_rate": 0.0001,
+      "loss": 2.9827,
+      "loss/crossentropy": 2.6424944162368775,
+      "loss/hidden": 2.66875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2500556200742722,
+      "loss/reg": 0.0,
+      "step": 32630
+    },
+    {
+      "epoch": 0.21473684210526317,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 6.259407552083333,
+      "learning_rate": 0.0001,
+      "loss": 3.0884,
+      "loss/crossentropy": 2.5622973680496215,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23999259918928145,
+      "loss/reg": 0.0,
+      "step": 32640
+    },
+    {
+      "epoch": 0.21480263157894736,
+      "grad_norm": 6.65625,
+      "grad_norm_var": 1.3544270833333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0648,
+      "loss/crossentropy": 2.322819399833679,
+      "loss/hidden": 2.89375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.288960862159729,
+      "loss/reg": 0.0,
+      "step": 32650
+    },
+    {
+      "epoch": 0.21486842105263157,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 1.3127237955729167,
+      "learning_rate": 0.0001,
+      "loss": 3.0818,
+      "loss/crossentropy": 2.179678440093994,
+      "loss/hidden": 2.89375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31069366484880445,
+      "loss/reg": 0.0,
+      "step": 32660
+    },
+    {
+      "epoch": 0.21493421052631578,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.10913798014322916,
+      "learning_rate": 0.0001,
+      "loss": 3.0115,
+      "loss/crossentropy": 2.2929873704910277,
+      "loss/hidden": 2.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24380201399326323,
+      "loss/reg": 0.0,
+      "step": 32670
+    },
+    {
+      "epoch": 0.215,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.05563151041666667,
+      "learning_rate": 0.0001,
+      "loss": 3.017,
+      "loss/crossentropy": 1.9826122522354126,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18868125975131989,
+      "loss/reg": 0.0,
+      "step": 32680
+    },
+    {
+      "epoch": 0.2150657894736842,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.08579813639322917,
+      "learning_rate": 0.0001,
+      "loss": 2.9559,
+      "loss/crossentropy": 2.4682058572769163,
+      "loss/hidden": 2.6921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22441071420907974,
+      "loss/reg": 0.0,
+      "step": 32690
+    },
+    {
+      "epoch": 0.21513157894736842,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.07283503214518229,
+      "learning_rate": 0.0001,
+      "loss": 2.9565,
+      "loss/crossentropy": 2.255161941051483,
+      "loss/hidden": 2.9,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2831642434000969,
+      "loss/reg": 0.0,
+      "step": 32700
+    },
+    {
+      "epoch": 0.21519736842105264,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.1724273681640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0657,
+      "loss/crossentropy": 2.233939862251282,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21798643767833709,
+      "loss/reg": 0.0,
+      "step": 32710
+    },
+    {
+      "epoch": 0.21526315789473685,
+      "grad_norm": 3.1875,
+      "grad_norm_var": 1.3363352457682292,
+      "learning_rate": 0.0001,
+      "loss": 3.0325,
+      "loss/crossentropy": 2.5652535796165465,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2587194755673409,
+      "loss/reg": 0.0,
+      "step": 32720
+    },
+    {
+      "epoch": 0.21532894736842106,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.130419921875,
+      "learning_rate": 0.0001,
+      "loss": 3.0003,
+      "loss/crossentropy": 2.1636557817459106,
+      "loss/hidden": 2.703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20836876183748246,
+      "loss/reg": 0.0,
+      "step": 32730
+    },
+    {
+      "epoch": 0.21539473684210525,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.035302734375,
+      "learning_rate": 0.0001,
+      "loss": 2.9321,
+      "loss/crossentropy": 2.4362986207008364,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24660978615283966,
+      "loss/reg": 0.0,
+      "step": 32740
+    },
+    {
+      "epoch": 0.21546052631578946,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.035105133056640626,
+      "learning_rate": 0.0001,
+      "loss": 2.9716,
+      "loss/crossentropy": 2.255325746536255,
+      "loss/hidden": 2.871875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24538252800703048,
+      "loss/reg": 0.0,
+      "step": 32750
+    },
+    {
+      "epoch": 0.21552631578947368,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.09314676920572916,
+      "learning_rate": 0.0001,
+      "loss": 3.0161,
+      "loss/crossentropy": 2.329508912563324,
+      "loss/hidden": 2.6359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21971380114555358,
+      "loss/reg": 0.0,
+      "step": 32760
+    },
+    {
+      "epoch": 0.2155921052631579,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.20466206868489584,
+      "learning_rate": 0.0001,
+      "loss": 2.963,
+      "loss/crossentropy": 2.262472677230835,
+      "loss/hidden": 2.7,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20472683310508727,
+      "loss/reg": 0.0,
+      "step": 32770
+    },
+    {
+      "epoch": 0.2156578947368421,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.25344009399414064,
+      "learning_rate": 0.0001,
+      "loss": 3.0138,
+      "loss/crossentropy": 2.3205514192581176,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23903629779815674,
+      "loss/reg": 0.0,
+      "step": 32780
+    },
+    {
+      "epoch": 0.21572368421052632,
+      "grad_norm": 3.09375,
+      "grad_norm_var": 2.3024370829264322,
+      "learning_rate": 0.0001,
+      "loss": 3.0763,
+      "loss/crossentropy": 2.315890896320343,
+      "loss/hidden": 2.85625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24931414425373077,
+      "loss/reg": 0.0,
+      "step": 32790
+    },
+    {
+      "epoch": 0.21578947368421053,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.34871317545572916,
+      "learning_rate": 0.0001,
+      "loss": 3.0259,
+      "loss/crossentropy": 2.2797498226165773,
+      "loss/hidden": 2.6109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20434827357530594,
+      "loss/reg": 0.0,
+      "step": 32800
+    },
+    {
+      "epoch": 0.21585526315789474,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.021117146809895834,
+      "learning_rate": 0.0001,
+      "loss": 2.9999,
+      "loss/crossentropy": 2.095989489555359,
+      "loss/hidden": 3.10625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2900803714990616,
+      "loss/reg": 0.0,
+      "step": 32810
+    },
+    {
+      "epoch": 0.21592105263157896,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.06117121378580729,
+      "learning_rate": 0.0001,
+      "loss": 3.03,
+      "loss/crossentropy": 2.325262129306793,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28796064853668213,
+      "loss/reg": 0.0,
+      "step": 32820
+    },
+    {
+      "epoch": 0.21598684210526317,
+      "grad_norm": 4.96875,
+      "grad_norm_var": 0.48582356770833335,
+      "learning_rate": 0.0001,
+      "loss": 2.9917,
+      "loss/crossentropy": 2.521669340133667,
+      "loss/hidden": 2.7078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2139764204621315,
+      "loss/reg": 0.0,
+      "step": 32830
+    },
+    {
+      "epoch": 0.21605263157894736,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.48948567708333335,
+      "learning_rate": 0.0001,
+      "loss": 2.9862,
+      "loss/crossentropy": 2.479197156429291,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22702316492795943,
+      "loss/reg": 0.0,
+      "step": 32840
+    },
+    {
+      "epoch": 0.21611842105263157,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.04804280598958333,
+      "learning_rate": 0.0001,
+      "loss": 2.9695,
+      "loss/crossentropy": 1.8314170956611633,
+      "loss/hidden": 2.66875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19034175872802733,
+      "loss/reg": 0.0,
+      "step": 32850
+    },
+    {
+      "epoch": 0.21618421052631578,
+      "grad_norm": 3.15625,
+      "grad_norm_var": 0.12819798787434897,
+      "learning_rate": 0.0001,
+      "loss": 2.9729,
+      "loss/crossentropy": 2.3301822662353517,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23585740774869918,
+      "loss/reg": 0.0,
+      "step": 32860
+    },
+    {
+      "epoch": 0.21625,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 3.20617589974401e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0646,
+      "loss/crossentropy": 2.3723347544670106,
+      "loss/hidden": 2.725,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22372013479471206,
+      "loss/reg": 0.0,
+      "step": 32870
+    },
+    {
+      "epoch": 0.2163157894736842,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 3.2061758995523174e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0732,
+      "loss/crossentropy": 2.337008368968964,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24527821093797683,
+      "loss/reg": 0.0,
+      "step": 32880
+    },
+    {
+      "epoch": 0.21638157894736842,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.18332926432291666,
+      "learning_rate": 0.0001,
+      "loss": 3.029,
+      "loss/crossentropy": 2.056360971927643,
+      "loss/hidden": 2.678125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20862108021974562,
+      "loss/reg": 0.0,
+      "step": 32890
+    },
+    {
+      "epoch": 0.21644736842105264,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.2578277587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0287,
+      "loss/crossentropy": 2.425034213066101,
+      "loss/hidden": 2.6140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22822874337434768,
+      "loss/reg": 0.0,
+      "step": 32900
+    },
+    {
+      "epoch": 0.21651315789473685,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.1828125,
+      "learning_rate": 0.0001,
+      "loss": 3.046,
+      "loss/crossentropy": 2.0453652262687685,
+      "loss/hidden": 3.0015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2606233850121498,
+      "loss/reg": 0.0,
+      "step": 32910
+    },
+    {
+      "epoch": 0.21657894736842107,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.1352203369140625,
+      "learning_rate": 0.0001,
+      "loss": 2.992,
+      "loss/crossentropy": 2.245816957950592,
+      "loss/hidden": 2.7984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2316955327987671,
+      "loss/reg": 0.0,
+      "step": 32920
+    },
+    {
+      "epoch": 0.21664473684210525,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.11937255859375,
+      "learning_rate": 0.0001,
+      "loss": 2.8992,
+      "loss/crossentropy": 2.3973164916038514,
+      "loss/hidden": 2.6671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21730958074331283,
+      "loss/reg": 0.0,
+      "step": 32930
+    },
+    {
+      "epoch": 0.21671052631578946,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.16926167805989584,
+      "learning_rate": 0.0001,
+      "loss": 3.076,
+      "loss/crossentropy": 2.3131038188934325,
+      "loss/hidden": 3.0328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2888943269848824,
+      "loss/reg": 0.0,
+      "step": 32940
+    },
+    {
+      "epoch": 0.21677631578947368,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.1410552978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0965,
+      "loss/crossentropy": 2.441923999786377,
+      "loss/hidden": 2.7625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2226712241768837,
+      "loss/reg": 0.0,
+      "step": 32950
+    },
+    {
+      "epoch": 0.2168421052631579,
+      "grad_norm": 5.6875,
+      "grad_norm_var": 0.6874338785807291,
+      "learning_rate": 0.0001,
+      "loss": 3.0441,
+      "loss/crossentropy": 2.0053435921669007,
+      "loss/hidden": 2.71875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20919274985790254,
+      "loss/reg": 0.0,
+      "step": 32960
+    },
+    {
+      "epoch": 0.2169078947368421,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.7307291666666667,
+      "learning_rate": 0.0001,
+      "loss": 3.0422,
+      "loss/crossentropy": 2.2273124933242796,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2289419651031494,
+      "loss/reg": 0.0,
+      "step": 32970
+    },
+    {
+      "epoch": 0.21697368421052632,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.0592681884765625,
+      "learning_rate": 0.0001,
+      "loss": 2.972,
+      "loss/crossentropy": 2.315923011302948,
+      "loss/hidden": 2.603125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19858411103487014,
+      "loss/reg": 0.0,
+      "step": 32980
+    },
+    {
+      "epoch": 0.21703947368421053,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.06795145670572916,
+      "learning_rate": 0.0001,
+      "loss": 3.1097,
+      "loss/crossentropy": 2.0356208801269533,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18648817762732506,
+      "loss/reg": 0.0,
+      "step": 32990
+    },
+    {
+      "epoch": 0.21710526315789475,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.087548828125,
+      "learning_rate": 0.0001,
+      "loss": 3.1037,
+      "loss/crossentropy": 2.4323093771934508,
+      "loss/hidden": 3.028125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2574485644698143,
+      "loss/reg": 0.0,
+      "step": 33000
+    },
+    {
+      "epoch": 0.21717105263157896,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.08557535807291666,
+      "learning_rate": 0.0001,
+      "loss": 2.9955,
+      "loss/crossentropy": 2.2170889139175416,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23443114012479782,
+      "loss/reg": 0.0,
+      "step": 33010
+    },
+    {
+      "epoch": 0.21723684210526314,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.0577056884765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0261,
+      "loss/crossentropy": 2.2761878967285156,
+      "loss/hidden": 2.85,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23024188429117204,
+      "loss/reg": 0.0,
+      "step": 33020
+    },
+    {
+      "epoch": 0.21730263157894736,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.041844685872395836,
+      "learning_rate": 0.0001,
+      "loss": 3.0025,
+      "loss/crossentropy": 2.281536602973938,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2611322790384293,
+      "loss/reg": 0.0,
+      "step": 33030
+    },
+    {
+      "epoch": 0.21736842105263157,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 2.2070302327473956,
+      "learning_rate": 0.0001,
+      "loss": 3.0541,
+      "loss/crossentropy": 2.285531198978424,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27087974548339844,
+      "loss/reg": 0.0,
+      "step": 33040
+    },
+    {
+      "epoch": 0.21743421052631579,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 2.3124013264973957,
+      "learning_rate": 0.0001,
+      "loss": 3.0046,
+      "loss/crossentropy": 2.1499021500349045,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2288394898176193,
+      "loss/reg": 0.0,
+      "step": 33050
+    },
+    {
+      "epoch": 0.2175,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.185888671875,
+      "learning_rate": 0.0001,
+      "loss": 3.01,
+      "loss/crossentropy": 2.2262683510780334,
+      "loss/hidden": 2.71875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21322729140520097,
+      "loss/reg": 0.0,
+      "step": 33060
+    },
+    {
+      "epoch": 0.2175657894736842,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.19086278279622396,
+      "learning_rate": 0.0001,
+      "loss": 2.9547,
+      "loss/crossentropy": 2.512454855442047,
+      "loss/hidden": 2.6953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22262426763772963,
+      "loss/reg": 0.0,
+      "step": 33070
+    },
+    {
+      "epoch": 0.21763157894736843,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.06533584594726563,
+      "learning_rate": 0.0001,
+      "loss": 3.0191,
+      "loss/crossentropy": 2.4302139401435854,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21147249341011048,
+      "loss/reg": 0.0,
+      "step": 33080
+    },
+    {
+      "epoch": 0.21769736842105264,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.05804036458333333,
+      "learning_rate": 0.0001,
+      "loss": 2.9211,
+      "loss/crossentropy": 2.288132381439209,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20329886302351952,
+      "loss/reg": 0.0,
+      "step": 33090
+    },
+    {
+      "epoch": 0.21776315789473685,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.07825419108072916,
+      "learning_rate": 0.0001,
+      "loss": 2.9798,
+      "loss/crossentropy": 2.4942794919013975,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22770892977714538,
+      "loss/reg": 0.0,
+      "step": 33100
+    },
+    {
+      "epoch": 0.21782894736842107,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.46223551432291665,
+      "learning_rate": 0.0001,
+      "loss": 2.998,
+      "loss/crossentropy": 2.496278202533722,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23407908529043198,
+      "loss/reg": 0.0,
+      "step": 33110
+    },
+    {
+      "epoch": 0.21789473684210525,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.1009674072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0352,
+      "loss/crossentropy": 2.0836499094963075,
+      "loss/hidden": 2.740625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22313288152217864,
+      "loss/reg": 0.0,
+      "step": 33120
+    },
+    {
+      "epoch": 0.21796052631578947,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.0770660400390625,
+      "learning_rate": 0.0001,
+      "loss": 2.982,
+      "loss/crossentropy": 2.413359189033508,
+      "loss/hidden": 2.828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23794451206922532,
+      "loss/reg": 0.0,
+      "step": 33130
+    },
+    {
+      "epoch": 0.21802631578947368,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.17366536458333334,
+      "learning_rate": 0.0001,
+      "loss": 2.9588,
+      "loss/crossentropy": 2.390314495563507,
+      "loss/hidden": 2.621875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20836979001760483,
+      "loss/reg": 0.0,
+      "step": 33140
+    },
+    {
+      "epoch": 0.2180921052631579,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.1214263916015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0312,
+      "loss/crossentropy": 2.1217519342899323,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22108341604471207,
+      "loss/reg": 0.0,
+      "step": 33150
+    },
+    {
+      "epoch": 0.2181578947368421,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.023981730143229168,
+      "learning_rate": 0.0001,
+      "loss": 2.9862,
+      "loss/crossentropy": 2.3113851308822633,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27378889471292495,
+      "loss/reg": 0.0,
+      "step": 33160
+    },
+    {
+      "epoch": 0.21822368421052632,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.07545166015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0085,
+      "loss/crossentropy": 2.366367816925049,
+      "loss/hidden": 2.68125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20626646727323533,
+      "loss/reg": 0.0,
+      "step": 33170
+    },
+    {
+      "epoch": 0.21828947368421053,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.08708394368489583,
+      "learning_rate": 0.0001,
+      "loss": 2.9855,
+      "loss/crossentropy": 2.2448646426200867,
+      "loss/hidden": 2.8296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2878586873412132,
+      "loss/reg": 0.0,
+      "step": 33180
+    },
+    {
+      "epoch": 0.21835526315789475,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.0867828369140625,
+      "learning_rate": 0.0001,
+      "loss": 2.9798,
+      "loss/crossentropy": 2.2996390104293822,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22233397513628006,
+      "loss/reg": 0.0,
+      "step": 33190
+    },
+    {
+      "epoch": 0.21842105263157896,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.12769266764322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0705,
+      "loss/crossentropy": 2.22303249835968,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32309592962265016,
+      "loss/reg": 0.0,
+      "step": 33200
+    },
+    {
+      "epoch": 0.21848684210526315,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.07171122233072917,
+      "learning_rate": 0.0001,
+      "loss": 3.0575,
+      "loss/crossentropy": 2.455464780330658,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2447066068649292,
+      "loss/reg": 0.0,
+      "step": 33210
+    },
+    {
+      "epoch": 0.21855263157894736,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.13964436848958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0108,
+      "loss/crossentropy": 2.4379093527793883,
+      "loss/hidden": 2.653125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20427114516496658,
+      "loss/reg": 0.0,
+      "step": 33220
+    },
+    {
+      "epoch": 0.21861842105263157,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.14078776041666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0186,
+      "loss/crossentropy": 2.3393636345863342,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24139494746923446,
+      "loss/reg": 0.0,
+      "step": 33230
+    },
+    {
+      "epoch": 0.2186842105263158,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.017975870768229166,
+      "learning_rate": 0.0001,
+      "loss": 3.0254,
+      "loss/crossentropy": 2.313693457841873,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25786808133125305,
+      "loss/reg": 0.0,
+      "step": 33240
+    },
+    {
+      "epoch": 0.21875,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.048111979166666666,
+      "learning_rate": 0.0001,
+      "loss": 2.9975,
+      "loss/crossentropy": 2.3757463097572327,
+      "loss/hidden": 2.6921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21639765352010726,
+      "loss/reg": 0.0,
+      "step": 33250
+    },
+    {
+      "epoch": 0.2188157894736842,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.07258707682291667,
+      "learning_rate": 0.0001,
+      "loss": 2.9429,
+      "loss/crossentropy": 2.540412497520447,
+      "loss/hidden": 2.565625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1985073819756508,
+      "loss/reg": 0.0,
+      "step": 33260
+    },
+    {
+      "epoch": 0.21888157894736843,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.03837788899739583,
+      "learning_rate": 0.0001,
+      "loss": 3.0685,
+      "loss/crossentropy": 2.373577296733856,
+      "loss/hidden": 2.790625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2263040378689766,
+      "loss/reg": 0.0,
+      "step": 33270
+    },
+    {
+      "epoch": 0.21894736842105264,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.029686482747395833,
+      "learning_rate": 0.0001,
+      "loss": 2.9608,
+      "loss/crossentropy": 2.2283737421035767,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23245082497596742,
+      "loss/reg": 0.0,
+      "step": 33280
+    },
+    {
+      "epoch": 0.21901315789473685,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.09920145670572916,
+      "learning_rate": 0.0001,
+      "loss": 3.0048,
+      "loss/crossentropy": 2.2985653638839723,
+      "loss/hidden": 2.871875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27912328094244004,
+      "loss/reg": 0.0,
+      "step": 33290
+    },
+    {
+      "epoch": 0.21907894736842104,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.10875651041666666,
+      "learning_rate": 0.0001,
+      "loss": 3.005,
+      "loss/crossentropy": 2.342892110347748,
+      "loss/hidden": 2.8671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2504061296582222,
+      "loss/reg": 0.0,
+      "step": 33300
+    },
+    {
+      "epoch": 0.21914473684210525,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.04578348795572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0108,
+      "loss/crossentropy": 2.337607777118683,
+      "loss/hidden": 2.6703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20644718632102013,
+      "loss/reg": 0.0,
+      "step": 33310
+    },
+    {
+      "epoch": 0.21921052631578947,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.08603108723958333,
+      "learning_rate": 0.0001,
+      "loss": 2.9696,
+      "loss/crossentropy": 2.080331861972809,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27878897339105607,
+      "loss/reg": 0.0,
+      "step": 33320
+    },
+    {
+      "epoch": 0.21927631578947368,
+      "grad_norm": 1.96875,
+      "grad_norm_var": 0.17981363932291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0076,
+      "loss/crossentropy": 2.3555988073349,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24901487827301025,
+      "loss/reg": 0.0,
+      "step": 33330
+    },
+    {
+      "epoch": 0.2193421052631579,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.17130533854166666,
+      "learning_rate": 0.0001,
+      "loss": 3.0368,
+      "loss/crossentropy": 2.223631227016449,
+      "loss/hidden": 2.696875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19968705028295516,
+      "loss/reg": 0.0,
+      "step": 33340
+    },
+    {
+      "epoch": 0.2194078947368421,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.22228190104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0191,
+      "loss/crossentropy": 2.108024871349335,
+      "loss/hidden": 2.5875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20175025016069412,
+      "loss/reg": 0.0,
+      "step": 33350
+    },
+    {
+      "epoch": 0.21947368421052632,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.2382720947265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0196,
+      "loss/crossentropy": 2.2824084401130675,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20648818761110305,
+      "loss/reg": 0.0,
+      "step": 33360
+    },
+    {
+      "epoch": 0.21953947368421053,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.06386311848958333,
+      "learning_rate": 0.0001,
+      "loss": 2.983,
+      "loss/crossentropy": 2.193799388408661,
+      "loss/hidden": 2.6,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20398449748754502,
+      "loss/reg": 0.0,
+      "step": 33370
+    },
+    {
+      "epoch": 0.21960526315789475,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.21061375935872395,
+      "learning_rate": 0.0001,
+      "loss": 3.0511,
+      "loss/crossentropy": 2.4952160239219667,
+      "loss/hidden": 2.7390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21543749868869783,
+      "loss/reg": 0.0,
+      "step": 33380
+    },
+    {
+      "epoch": 0.21967105263157893,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.6498410542805989,
+      "learning_rate": 0.0001,
+      "loss": 3.0987,
+      "loss/crossentropy": 2.4116694569587707,
+      "loss/hidden": 2.921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2724195793271065,
+      "loss/reg": 0.0,
+      "step": 33390
+    },
+    {
+      "epoch": 0.21973684210526315,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.23567072550455728,
+      "learning_rate": 0.0001,
+      "loss": 2.9416,
+      "loss/crossentropy": 2.597557008266449,
+      "loss/hidden": 2.809375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2337816461920738,
+      "loss/reg": 0.0,
+      "step": 33400
+    },
+    {
+      "epoch": 0.21980263157894736,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.03090387980143229,
+      "learning_rate": 0.0001,
+      "loss": 2.9881,
+      "loss/crossentropy": 2.446978306770325,
+      "loss/hidden": 2.609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2176991730928421,
+      "loss/reg": 0.0,
+      "step": 33410
+    },
+    {
+      "epoch": 0.21986842105263157,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.0204742431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0409,
+      "loss/crossentropy": 2.0214170694351195,
+      "loss/hidden": 3.0140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23912209570407866,
+      "loss/reg": 0.0,
+      "step": 33420
+    },
+    {
+      "epoch": 0.2199342105263158,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.01568603515625,
+      "learning_rate": 0.0001,
+      "loss": 3.042,
+      "loss/crossentropy": 2.4408559799194336,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23146803677082062,
+      "loss/reg": 0.0,
+      "step": 33430
+    },
+    {
+      "epoch": 0.22,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.05896809895833333,
+      "learning_rate": 0.0001,
+      "loss": 3.0667,
+      "loss/crossentropy": 2.699678826332092,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.250095309317112,
+      "loss/reg": 0.0,
+      "step": 33440
+    },
+    {
+      "epoch": 0.22006578947368421,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.14205322265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0448,
+      "loss/crossentropy": 2.2925549387931823,
+      "loss/hidden": 2.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24470344185829163,
+      "loss/reg": 0.0,
+      "step": 33450
+    },
+    {
+      "epoch": 0.22013157894736843,
+      "grad_norm": 2.25,
+      "grad_norm_var": 3.588060506184896,
+      "learning_rate": 0.0001,
+      "loss": 3.1012,
+      "loss/crossentropy": 2.3695211172103883,
+      "loss/hidden": 2.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24156490415334703,
+      "loss/reg": 0.0,
+      "step": 33460
+    },
+    {
+      "epoch": 0.22019736842105264,
+      "grad_norm": 2382364672.0,
+      "grad_norm_var": 3.5472883849629094e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1654,
+      "loss/crossentropy": 2.488659930229187,
+      "loss/hidden": 2.7390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21741390377283096,
+      "loss/reg": 0.0,
+      "step": 33470
+    },
+    {
+      "epoch": 0.22026315789473686,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 3.547288386504622e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0377,
+      "loss/crossentropy": 2.2884308457374574,
+      "loss/hidden": 2.8984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.231295208632946,
+      "loss/reg": 0.0,
+      "step": 33480
+    },
+    {
+      "epoch": 0.22032894736842104,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.049658203125,
+      "learning_rate": 0.0001,
+      "loss": 3.06,
+      "loss/crossentropy": 2.163191545009613,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22781991362571716,
+      "loss/reg": 0.0,
+      "step": 33490
+    },
+    {
+      "epoch": 0.22039473684210525,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.08287353515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9897,
+      "loss/crossentropy": 2.2403807282447814,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23052917122840882,
+      "loss/reg": 0.0,
+      "step": 33500
+    },
+    {
+      "epoch": 0.22046052631578947,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.4391029357910156,
+      "learning_rate": 0.0001,
+      "loss": 2.9359,
+      "loss/crossentropy": 2.215052628517151,
+      "loss/hidden": 2.6265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.17493754550814627,
+      "loss/reg": 0.0,
+      "step": 33510
+    },
+    {
+      "epoch": 0.22052631578947368,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 1.1144650777180989,
+      "learning_rate": 0.0001,
+      "loss": 3.0608,
+      "loss/crossentropy": 2.0380016922950746,
+      "loss/hidden": 2.90625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2704147264361382,
+      "loss/reg": 0.0,
+      "step": 33520
+    },
+    {
+      "epoch": 0.2205921052631579,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 1.154686482747396,
+      "learning_rate": 0.0001,
+      "loss": 3.0527,
+      "loss/crossentropy": 2.5095550060272216,
+      "loss/hidden": 2.9453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2915314584970474,
+      "loss/reg": 0.0,
+      "step": 33530
+    },
+    {
+      "epoch": 0.2206578947368421,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.4706939697265625,
+      "learning_rate": 0.0001,
+      "loss": 2.9389,
+      "loss/crossentropy": 2.175918400287628,
+      "loss/hidden": 2.59375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19029978811740875,
+      "loss/reg": 0.0,
+      "step": 33540
+    },
+    {
+      "epoch": 0.22072368421052632,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.13570556640625,
+      "learning_rate": 0.0001,
+      "loss": 3.1049,
+      "loss/crossentropy": 2.1491673469543455,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28144592940807345,
+      "loss/reg": 0.0,
+      "step": 33550
+    },
+    {
+      "epoch": 0.22078947368421054,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.20730692545572918,
+      "learning_rate": 0.0001,
+      "loss": 3.0324,
+      "loss/crossentropy": 2.10293396115303,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20049710869789122,
+      "loss/reg": 0.0,
+      "step": 33560
+    },
+    {
+      "epoch": 0.22085526315789475,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.20046284993489583,
+      "learning_rate": 0.0001,
+      "loss": 3.082,
+      "loss/crossentropy": 2.4584061741828918,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.235392826795578,
+      "loss/reg": 0.0,
+      "step": 33570
+    },
+    {
+      "epoch": 0.22092105263157893,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.0829742431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0276,
+      "loss/crossentropy": 2.107621490955353,
+      "loss/hidden": 2.8390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2272868797183037,
+      "loss/reg": 0.0,
+      "step": 33580
+    },
+    {
+      "epoch": 0.22098684210526315,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.25580952962239584,
+      "learning_rate": 0.0001,
+      "loss": 3.0614,
+      "loss/crossentropy": 2.120447838306427,
+      "loss/hidden": 2.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19922292605042458,
+      "loss/reg": 0.0,
+      "step": 33590
+    },
+    {
+      "epoch": 0.22105263157894736,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.027180989583333332,
+      "learning_rate": 0.0001,
+      "loss": 2.9857,
+      "loss/crossentropy": 2.192478084564209,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2926119461655617,
+      "loss/reg": 0.0,
+      "step": 33600
+    },
+    {
+      "epoch": 0.22111842105263158,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.07001851399739584,
+      "learning_rate": 0.0001,
+      "loss": 3.0663,
+      "loss/crossentropy": 2.4031628251075743,
+      "loss/hidden": 2.6421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21948793083429335,
+      "loss/reg": 0.0,
+      "step": 33610
+    },
+    {
+      "epoch": 0.2211842105263158,
+      "grad_norm": 3.671875,
+      "grad_norm_var": 0.183837890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0731,
+      "loss/crossentropy": 2.2898489594459535,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2655272573232651,
+      "loss/reg": 0.0,
+      "step": 33620
+    },
+    {
+      "epoch": 0.22125,
+      "grad_norm": 2.25,
+      "grad_norm_var": 1.4440388997395834,
+      "learning_rate": 0.0001,
+      "loss": 3.0888,
+      "loss/crossentropy": 2.334256184101105,
+      "loss/hidden": 2.846875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20996386408805848,
+      "loss/reg": 0.0,
+      "step": 33630
+    },
+    {
+      "epoch": 0.22131578947368422,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.04494527180989583,
+      "learning_rate": 0.0001,
+      "loss": 3.0546,
+      "loss/crossentropy": 2.3563809156417848,
+      "loss/hidden": 3.0484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29970877766609194,
+      "loss/reg": 0.0,
+      "step": 33640
+    },
+    {
+      "epoch": 0.22138157894736843,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.040283203125,
+      "learning_rate": 0.0001,
+      "loss": 3.0316,
+      "loss/crossentropy": 2.469383955001831,
+      "loss/hidden": 2.765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22515997290611267,
+      "loss/reg": 0.0,
+      "step": 33650
+    },
+    {
+      "epoch": 0.22144736842105264,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.0654449462890625,
+      "learning_rate": 0.0001,
+      "loss": 3.1037,
+      "loss/crossentropy": 2.4161542892456054,
+      "loss/hidden": 2.9859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22671497017145156,
+      "loss/reg": 0.0,
+      "step": 33660
+    },
+    {
+      "epoch": 0.22151315789473683,
+      "grad_norm": 3.953125,
+      "grad_norm_var": 0.27838134765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0341,
+      "loss/crossentropy": 2.415839672088623,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2382939413189888,
+      "loss/reg": 0.0,
+      "step": 33670
+    },
+    {
+      "epoch": 0.22157894736842104,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.9553260803222656,
+      "learning_rate": 0.0001,
+      "loss": 3.0584,
+      "loss/crossentropy": 2.353093123435974,
+      "loss/hidden": 2.890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.276655937731266,
+      "loss/reg": 0.0,
+      "step": 33680
+    },
+    {
+      "epoch": 0.22164473684210526,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.8398048400878906,
+      "learning_rate": 0.0001,
+      "loss": 3.0327,
+      "loss/crossentropy": 2.2812642812728883,
+      "loss/hidden": 2.684375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20394409000873565,
+      "loss/reg": 0.0,
+      "step": 33690
+    },
+    {
+      "epoch": 0.22171052631578947,
+      "grad_norm": 1.859375,
+      "grad_norm_var": 0.10286051432291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0121,
+      "loss/crossentropy": 2.2089013338088987,
+      "loss/hidden": 2.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2560999572277069,
+      "loss/reg": 0.0,
+      "step": 33700
+    },
+    {
+      "epoch": 0.22177631578947368,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.224267578125,
+      "learning_rate": 0.0001,
+      "loss": 2.9909,
+      "loss/crossentropy": 2.579922378063202,
+      "loss/hidden": 2.75625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2395647093653679,
+      "loss/reg": 0.0,
+      "step": 33710
+    },
+    {
+      "epoch": 0.2218421052631579,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.2459307352701823,
+      "learning_rate": 0.0001,
+      "loss": 2.9947,
+      "loss/crossentropy": 2.269551432132721,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22979858070611953,
+      "loss/reg": 0.0,
+      "step": 33720
+    },
+    {
+      "epoch": 0.2219078947368421,
+      "grad_norm": 2.0,
+      "grad_norm_var": 0.0810455322265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0223,
+      "loss/crossentropy": 2.5746334075927733,
+      "loss/hidden": 2.675,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.237695574760437,
+      "loss/reg": 0.0,
+      "step": 33730
+    },
+    {
+      "epoch": 0.22197368421052632,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.0932281494140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0095,
+      "loss/crossentropy": 2.111787271499634,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21114924997091294,
+      "loss/reg": 0.0,
+      "step": 33740
+    },
+    {
+      "epoch": 0.22203947368421054,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.1626129150390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0218,
+      "loss/crossentropy": 2.392720115184784,
+      "loss/hidden": 2.6671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19515267461538316,
+      "loss/reg": 0.0,
+      "step": 33750
+    },
+    {
+      "epoch": 0.22210526315789475,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.049347941080729166,
+      "learning_rate": 0.0001,
+      "loss": 3.0357,
+      "loss/crossentropy": 2.4739153146743775,
+      "loss/hidden": 2.9296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3125626042485237,
+      "loss/reg": 0.0,
+      "step": 33760
+    },
+    {
+      "epoch": 0.22217105263157894,
+      "grad_norm": 1.9453125,
+      "grad_norm_var": 0.1904070536295573,
+      "learning_rate": 0.0001,
+      "loss": 2.974,
+      "loss/crossentropy": 2.290574276447296,
+      "loss/hidden": 2.8140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23448731899261474,
+      "loss/reg": 0.0,
+      "step": 33770
+    },
+    {
+      "epoch": 0.22223684210526315,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.13733495076497396,
+      "learning_rate": 0.0001,
+      "loss": 2.9673,
+      "loss/crossentropy": 2.317961239814758,
+      "loss/hidden": 2.6984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21905806362628938,
+      "loss/reg": 0.0,
+      "step": 33780
+    },
+    {
+      "epoch": 0.22230263157894736,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.16253255208333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0608,
+      "loss/crossentropy": 2.170513927936554,
+      "loss/hidden": 2.83125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20969858914613723,
+      "loss/reg": 0.0,
+      "step": 33790
+    },
+    {
+      "epoch": 0.22236842105263158,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.20142822265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0502,
+      "loss/crossentropy": 2.3547741293907167,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21735511124134063,
+      "loss/reg": 0.0,
+      "step": 33800
+    },
+    {
+      "epoch": 0.2224342105263158,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.62681884765625,
+      "learning_rate": 0.0001,
+      "loss": 2.9739,
+      "loss/crossentropy": 2.3395539283752442,
+      "loss/hidden": 2.653125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20967817306518555,
+      "loss/reg": 0.0,
+      "step": 33810
+    },
+    {
+      "epoch": 0.2225,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.0564117431640625,
+      "learning_rate": 0.0001,
+      "loss": 2.9871,
+      "loss/crossentropy": 2.294986367225647,
+      "loss/hidden": 2.815625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2147073432803154,
+      "loss/reg": 0.0,
+      "step": 33820
+    },
+    {
+      "epoch": 0.22256578947368422,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.07506103515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9679,
+      "loss/crossentropy": 2.257847762107849,
+      "loss/hidden": 2.9234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30166392475366594,
+      "loss/reg": 0.0,
+      "step": 33830
+    },
+    {
+      "epoch": 0.22263157894736843,
+      "grad_norm": 3.40625,
+      "grad_norm_var": 0.17619527180989583,
+      "learning_rate": 0.0001,
+      "loss": 2.9993,
+      "loss/crossentropy": 2.2673079133033753,
+      "loss/hidden": 2.646875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21112384274601936,
+      "loss/reg": 0.0,
+      "step": 33840
+    },
+    {
+      "epoch": 0.22269736842105264,
+      "grad_norm": 5.0,
+      "grad_norm_var": 0.5702311197916666,
+      "learning_rate": 0.0001,
+      "loss": 3.0953,
+      "loss/crossentropy": 2.632491409778595,
+      "loss/hidden": 2.753125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23320973068475723,
+      "loss/reg": 0.0,
+      "step": 33850
+    },
+    {
+      "epoch": 0.22276315789473683,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.7403310139973959,
+      "learning_rate": 0.0001,
+      "loss": 3.0625,
+      "loss/crossentropy": 2.331642270088196,
+      "loss/hidden": 2.88125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23803979456424712,
+      "loss/reg": 0.0,
+      "step": 33860
+    },
+    {
+      "epoch": 0.22282894736842104,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.041890462239583336,
+      "learning_rate": 0.0001,
+      "loss": 3.034,
+      "loss/crossentropy": 2.2647625207901,
+      "loss/hidden": 2.7203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20834072977304458,
+      "loss/reg": 0.0,
+      "step": 33870
+    },
+    {
+      "epoch": 0.22289473684210526,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.0492095947265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0249,
+      "loss/crossentropy": 2.259378707408905,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21118512228131295,
+      "loss/reg": 0.0,
+      "step": 33880
+    },
+    {
+      "epoch": 0.22296052631578947,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.10131810506184896,
+      "learning_rate": 0.0001,
+      "loss": 3.0156,
+      "loss/crossentropy": 2.31290363073349,
+      "loss/hidden": 2.63125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2051762267947197,
+      "loss/reg": 0.0,
+      "step": 33890
+    },
+    {
+      "epoch": 0.22302631578947368,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.11342544555664062,
+      "learning_rate": 0.0001,
+      "loss": 3.0365,
+      "loss/crossentropy": 1.8784321069717407,
+      "loss/hidden": 2.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21909460723400115,
+      "loss/reg": 0.0,
+      "step": 33900
+    },
+    {
+      "epoch": 0.2230921052631579,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.05671284993489583,
+      "learning_rate": 0.0001,
+      "loss": 3.016,
+      "loss/crossentropy": 2.3022819638252257,
+      "loss/hidden": 2.7265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22228346467018129,
+      "loss/reg": 0.0,
+      "step": 33910
+    },
+    {
+      "epoch": 0.2231578947368421,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.23657938639322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0624,
+      "loss/crossentropy": 2.335461437702179,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23024386763572693,
+      "loss/reg": 0.0,
+      "step": 33920
+    },
+    {
+      "epoch": 0.22322368421052632,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.18298238118489582,
+      "learning_rate": 0.0001,
+      "loss": 3.1342,
+      "loss/crossentropy": 2.149783802032471,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24487016648054122,
+      "loss/reg": 0.0,
+      "step": 33930
+    },
+    {
+      "epoch": 0.22328947368421054,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.07650731404622396,
+      "learning_rate": 0.0001,
+      "loss": 3.0341,
+      "loss/crossentropy": 2.248632514476776,
+      "loss/hidden": 2.846875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22311466187238693,
+      "loss/reg": 0.0,
+      "step": 33940
+    },
+    {
+      "epoch": 0.22335526315789472,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.0645416259765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1104,
+      "loss/crossentropy": 2.367274534702301,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2507220461964607,
+      "loss/reg": 0.0,
+      "step": 33950
+    },
+    {
+      "epoch": 0.22342105263157894,
+      "grad_norm": 2.90625,
+      "grad_norm_var": 0.05906473795572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0713,
+      "loss/crossentropy": 2.3802406549453736,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23559605777263642,
+      "loss/reg": 0.0,
+      "step": 33960
+    },
+    {
+      "epoch": 0.22348684210526315,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.045979817708333336,
+      "learning_rate": 0.0001,
+      "loss": 3.0398,
+      "loss/crossentropy": 2.283581781387329,
+      "loss/hidden": 2.8359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22127386182546616,
+      "loss/reg": 0.0,
+      "step": 33970
+    },
+    {
+      "epoch": 0.22355263157894736,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.0328277587890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0036,
+      "loss/crossentropy": 2.2841804146766664,
+      "loss/hidden": 2.9328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29590369313955306,
+      "loss/reg": 0.0,
+      "step": 33980
+    },
+    {
+      "epoch": 0.22361842105263158,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.0925201416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0846,
+      "loss/crossentropy": 2.1158400774002075,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21308634355664252,
+      "loss/reg": 0.0,
+      "step": 33990
+    },
+    {
+      "epoch": 0.2236842105263158,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.1506744384765625,
+      "learning_rate": 0.0001,
+      "loss": 3.081,
+      "loss/crossentropy": 2.3101623356342316,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22568342536687852,
+      "loss/reg": 0.0,
+      "step": 34000
+    },
+    {
+      "epoch": 0.22375,
+      "grad_norm": 3.28125,
+      "grad_norm_var": 0.21297200520833334,
+      "learning_rate": 0.0001,
+      "loss": 3.0695,
+      "loss/crossentropy": 2.493696868419647,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32419712096452713,
+      "loss/reg": 0.0,
+      "step": 34010
+    },
+    {
+      "epoch": 0.22381578947368422,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.470361328125,
+      "learning_rate": 0.0001,
+      "loss": 3.0831,
+      "loss/crossentropy": 2.181097960472107,
+      "loss/hidden": 2.828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21221623122692107,
+      "loss/reg": 0.0,
+      "step": 34020
+    },
+    {
+      "epoch": 0.22388157894736843,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.4900716145833333,
+      "learning_rate": 0.0001,
+      "loss": 2.9904,
+      "loss/crossentropy": 2.2486413717269897,
+      "loss/hidden": 2.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2214418590068817,
+      "loss/reg": 0.0,
+      "step": 34030
+    },
+    {
+      "epoch": 0.22394736842105264,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.026363118489583334,
+      "learning_rate": 0.0001,
+      "loss": 3.0049,
+      "loss/crossentropy": 2.058054494857788,
+      "loss/hidden": 2.678125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20110160410404204,
+      "loss/reg": 0.0,
+      "step": 34040
+    },
+    {
+      "epoch": 0.22401315789473683,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.8867828369140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0557,
+      "loss/crossentropy": 2.183249998092651,
+      "loss/hidden": 2.94375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3117926768958569,
+      "loss/reg": 0.0,
+      "step": 34050
+    },
+    {
+      "epoch": 0.22407894736842104,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.73414306640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0466,
+      "loss/crossentropy": 2.2317178010940553,
+      "loss/hidden": 2.8640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23523171246051788,
+      "loss/reg": 0.0,
+      "step": 34060
+    },
+    {
+      "epoch": 0.22414473684210526,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.11988525390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0361,
+      "loss/crossentropy": 2.1958236932754516,
+      "loss/hidden": 2.7984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2175719380378723,
+      "loss/reg": 0.0,
+      "step": 34070
+    },
+    {
+      "epoch": 0.22421052631578947,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.08212483723958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0053,
+      "loss/crossentropy": 2.3172508239746095,
+      "loss/hidden": 2.640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21442833691835403,
+      "loss/reg": 0.0,
+      "step": 34080
+    },
+    {
+      "epoch": 0.22427631578947368,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.13967666625976563,
+      "learning_rate": 0.0001,
+      "loss": 2.9913,
+      "loss/crossentropy": 2.244894874095917,
+      "loss/hidden": 2.6546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19675317481160165,
+      "loss/reg": 0.0,
+      "step": 34090
+    },
+    {
+      "epoch": 0.2243421052631579,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.16440200805664062,
+      "learning_rate": 0.0001,
+      "loss": 3.046,
+      "loss/crossentropy": 2.256273639202118,
+      "loss/hidden": 2.6890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22018496468663215,
+      "loss/reg": 0.0,
+      "step": 34100
+    },
+    {
+      "epoch": 0.2244078947368421,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.15823160807291667,
+      "learning_rate": 0.0001,
+      "loss": 2.9706,
+      "loss/crossentropy": 2.450822639465332,
+      "loss/hidden": 2.896875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.31932070925831796,
+      "loss/reg": 0.0,
+      "step": 34110
+    },
+    {
+      "epoch": 0.22447368421052633,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.212939453125,
+      "learning_rate": 0.0001,
+      "loss": 3.0805,
+      "loss/crossentropy": 2.363582265377045,
+      "loss/hidden": 3.0015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24354881346225737,
+      "loss/reg": 0.0,
+      "step": 34120
+    },
+    {
+      "epoch": 0.22453947368421054,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.07653706868489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0054,
+      "loss/crossentropy": 2.4329180479049684,
+      "loss/hidden": 2.809375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2980032041668892,
+      "loss/reg": 0.0,
+      "step": 34130
+    },
+    {
+      "epoch": 0.22460526315789472,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.14047012329101563,
+      "learning_rate": 0.0001,
+      "loss": 3.0115,
+      "loss/crossentropy": 2.339194095134735,
+      "loss/hidden": 2.7359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22101039290428162,
+      "loss/reg": 0.0,
+      "step": 34140
+    },
+    {
+      "epoch": 0.22467105263157894,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.14289957682291668,
+      "learning_rate": 0.0001,
+      "loss": 2.9833,
+      "loss/crossentropy": 2.1932947993278504,
+      "loss/hidden": 2.9234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2560951545834541,
+      "loss/reg": 0.0,
+      "step": 34150
+    },
+    {
+      "epoch": 0.22473684210526315,
+      "grad_norm": 2.0,
+      "grad_norm_var": 0.11972249348958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0913,
+      "loss/crossentropy": 2.341690111160278,
+      "loss/hidden": 2.6296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21807242110371589,
+      "loss/reg": 0.0,
+      "step": 34160
+    },
+    {
+      "epoch": 0.22480263157894737,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.039937337239583336,
+      "learning_rate": 0.0001,
+      "loss": 3.0886,
+      "loss/crossentropy": 1.9180511385202408,
+      "loss/hidden": 2.9625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2749033223837614,
+      "loss/reg": 0.0,
+      "step": 34170
+    },
+    {
+      "epoch": 0.22486842105263158,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.0544586181640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0813,
+      "loss/crossentropy": 2.0566360354423523,
+      "loss/hidden": 2.9953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22760727405548095,
+      "loss/reg": 0.0,
+      "step": 34180
+    },
+    {
+      "epoch": 0.2249342105263158,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.0463043212890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0512,
+      "loss/crossentropy": 2.291218078136444,
+      "loss/hidden": 2.7375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23884439915418626,
+      "loss/reg": 0.0,
+      "step": 34190
+    },
+    {
+      "epoch": 0.225,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.04059956868489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0016,
+      "loss/crossentropy": 2.409184718132019,
+      "loss/hidden": 2.646875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22745371460914612,
+      "loss/reg": 0.0,
+      "step": 34200
+    },
+    {
+      "epoch": 0.22506578947368422,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.028944651285807293,
+      "learning_rate": 0.0001,
+      "loss": 3.0011,
+      "loss/crossentropy": 2.541308951377869,
+      "loss/hidden": 2.5921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19628863781690598,
+      "loss/reg": 0.0,
+      "step": 34210
+    },
+    {
+      "epoch": 0.22513157894736843,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.080078125,
+      "learning_rate": 0.0001,
+      "loss": 3.0556,
+      "loss/crossentropy": 1.890171855688095,
+      "loss/hidden": 2.975,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23669045567512512,
+      "loss/reg": 0.0,
+      "step": 34220
+    },
+    {
+      "epoch": 0.22519736842105262,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.031754557291666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0229,
+      "loss/crossentropy": 1.9730781435966491,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2429889589548111,
+      "loss/reg": 0.0,
+      "step": 34230
+    },
+    {
+      "epoch": 0.22526315789473683,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.12756754557291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0908,
+      "loss/crossentropy": 2.430511474609375,
+      "loss/hidden": 2.928125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26906117498874665,
+      "loss/reg": 0.0,
+      "step": 34240
+    },
+    {
+      "epoch": 0.22532894736842105,
+      "grad_norm": 3.5625,
+      "grad_norm_var": 0.36015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1105,
+      "loss/crossentropy": 2.387774920463562,
+      "loss/hidden": 2.9265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23492622524499893,
+      "loss/reg": 0.0,
+      "step": 34250
+    },
+    {
+      "epoch": 0.22539473684210526,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.5035804748535156,
+      "learning_rate": 0.0001,
+      "loss": 2.9848,
+      "loss/crossentropy": 2.3150025963783265,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22484272122383117,
+      "loss/reg": 0.0,
+      "step": 34260
+    },
+    {
+      "epoch": 0.22546052631578947,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.47118504842122394,
+      "learning_rate": 0.0001,
+      "loss": 2.9996,
+      "loss/crossentropy": 2.1695404648780823,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20138196349143983,
+      "loss/reg": 0.0,
+      "step": 34270
+    },
+    {
+      "epoch": 0.22552631578947369,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.055063629150390626,
+      "learning_rate": 0.0001,
+      "loss": 2.9955,
+      "loss/crossentropy": 2.348333418369293,
+      "loss/hidden": 2.8265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2416067436337471,
+      "loss/reg": 0.0,
+      "step": 34280
+    },
+    {
+      "epoch": 0.2255921052631579,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.1240631103515625,
+      "learning_rate": 0.0001,
+      "loss": 3.051,
+      "loss/crossentropy": 2.1614134192466734,
+      "loss/hidden": 2.6953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2166384145617485,
+      "loss/reg": 0.0,
+      "step": 34290
+    },
+    {
+      "epoch": 0.2256578947368421,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.18293355305989584,
+      "learning_rate": 0.0001,
+      "loss": 2.9882,
+      "loss/crossentropy": 2.1018754601478578,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19822993129491806,
+      "loss/reg": 0.0,
+      "step": 34300
+    },
+    {
+      "epoch": 0.22572368421052633,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.10800374348958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1428,
+      "loss/crossentropy": 2.4500314474105833,
+      "loss/hidden": 2.990625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3025324195623398,
+      "loss/reg": 0.0,
+      "step": 34310
+    },
+    {
+      "epoch": 0.22578947368421054,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.0825592041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0246,
+      "loss/crossentropy": 2.1041884064674377,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22658382654190062,
+      "loss/reg": 0.0,
+      "step": 34320
+    },
+    {
+      "epoch": 0.22585526315789473,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.07895406087239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0125,
+      "loss/crossentropy": 2.397817623615265,
+      "loss/hidden": 3.00625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28272619247436526,
+      "loss/reg": 0.0,
+      "step": 34330
+    },
+    {
+      "epoch": 0.22592105263157894,
+      "grad_norm": 1.953125,
+      "grad_norm_var": 0.08463134765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0419,
+      "loss/crossentropy": 2.0744609951972963,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27017629742622373,
+      "loss/reg": 0.0,
+      "step": 34340
+    },
+    {
+      "epoch": 0.22598684210526315,
+      "grad_norm": 3.171875,
+      "grad_norm_var": 0.10567118326822916,
+      "learning_rate": 0.0001,
+      "loss": 3.054,
+      "loss/crossentropy": 2.5243443489074706,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22290815711021422,
+      "loss/reg": 0.0,
+      "step": 34350
+    },
+    {
+      "epoch": 0.22605263157894737,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.20874735514322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0311,
+      "loss/crossentropy": 2.291029953956604,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23554186373949051,
+      "loss/reg": 0.0,
+      "step": 34360
+    },
+    {
+      "epoch": 0.22611842105263158,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.17711588541666667,
+      "learning_rate": 0.0001,
+      "loss": 3.0285,
+      "loss/crossentropy": 2.4715088486671446,
+      "loss/hidden": 2.7453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2226739391684532,
+      "loss/reg": 0.0,
+      "step": 34370
+    },
+    {
+      "epoch": 0.2261842105263158,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.07121988932291666,
+      "learning_rate": 0.0001,
+      "loss": 2.9697,
+      "loss/crossentropy": 2.226884996891022,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19653952047228812,
+      "loss/reg": 0.0,
+      "step": 34380
+    },
+    {
+      "epoch": 0.22625,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.11801656087239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0042,
+      "loss/crossentropy": 2.6288553953170775,
+      "loss/hidden": 2.7875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22481338530778885,
+      "loss/reg": 0.0,
+      "step": 34390
+    },
+    {
+      "epoch": 0.22631578947368422,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.056029256184895834,
+      "learning_rate": 0.0001,
+      "loss": 2.9881,
+      "loss/crossentropy": 2.086442506313324,
+      "loss/hidden": 2.6640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19545092582702636,
+      "loss/reg": 0.0,
+      "step": 34400
+    },
+    {
+      "epoch": 0.22638157894736843,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.07093098958333334,
+      "learning_rate": 0.0001,
+      "loss": 2.9785,
+      "loss/crossentropy": 2.1755173921585085,
+      "loss/hidden": 2.78125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2160940498113632,
+      "loss/reg": 0.0,
+      "step": 34410
+    },
+    {
+      "epoch": 0.22644736842105262,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.5419667561848959,
+      "learning_rate": 0.0001,
+      "loss": 3.148,
+      "loss/crossentropy": 2.306460678577423,
+      "loss/hidden": 2.9765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23442021161317825,
+      "loss/reg": 0.0,
+      "step": 34420
+    },
+    {
+      "epoch": 0.22651315789473683,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.509814453125,
+      "learning_rate": 0.0001,
+      "loss": 3.144,
+      "loss/crossentropy": 2.4273258805274964,
+      "loss/hidden": 2.9859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2834622025489807,
+      "loss/reg": 0.0,
+      "step": 34430
+    },
+    {
+      "epoch": 0.22657894736842105,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.20230712890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0404,
+      "loss/crossentropy": 2.382106697559357,
+      "loss/hidden": 2.840625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2566205784678459,
+      "loss/reg": 0.0,
+      "step": 34440
+    },
+    {
+      "epoch": 0.22664473684210526,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 1.1934967041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.1251,
+      "loss/crossentropy": 2.4071131587028503,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25932526737451556,
+      "loss/reg": 0.0,
+      "step": 34450
+    },
+    {
+      "epoch": 0.22671052631578947,
+      "grad_norm": 3.65625,
+      "grad_norm_var": 0.17838109334309896,
+      "learning_rate": 0.0001,
+      "loss": 3.0335,
+      "loss/crossentropy": 2.3701361656188964,
+      "loss/hidden": 2.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2191499412059784,
+      "loss/reg": 0.0,
+      "step": 34460
+    },
+    {
+      "epoch": 0.2267763157894737,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.1755633036295573,
+      "learning_rate": 0.0001,
+      "loss": 3.1017,
+      "loss/crossentropy": 1.7690342009067535,
+      "loss/hidden": 2.83125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19500833228230477,
+      "loss/reg": 0.0,
+      "step": 34470
+    },
+    {
+      "epoch": 0.2268421052631579,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.36656494140625,
+      "learning_rate": 0.0001,
+      "loss": 3.014,
+      "loss/crossentropy": 2.1445329546928407,
+      "loss/hidden": 2.6640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.192388217151165,
+      "loss/reg": 0.0,
+      "step": 34480
+    },
+    {
+      "epoch": 0.22690789473684211,
+      "grad_norm": 3.65625,
+      "grad_norm_var": 0.5748331705729167,
+      "learning_rate": 0.0001,
+      "loss": 3.156,
+      "loss/crossentropy": 2.152642047405243,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23683252632617952,
+      "loss/reg": 0.0,
+      "step": 34490
+    },
+    {
+      "epoch": 0.22697368421052633,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.1991119384765625,
+      "learning_rate": 0.0001,
+      "loss": 3.1698,
+      "loss/crossentropy": 2.343445897102356,
+      "loss/hidden": 2.8265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25687034577131274,
+      "loss/reg": 0.0,
+      "step": 34500
+    },
+    {
+      "epoch": 0.2270394736842105,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.1173828125,
+      "learning_rate": 0.0001,
+      "loss": 3.0052,
+      "loss/crossentropy": 2.352896881103516,
+      "loss/hidden": 2.7578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27792870104312895,
+      "loss/reg": 0.0,
+      "step": 34510
+    },
+    {
+      "epoch": 0.22710526315789473,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.05995992024739583,
+      "learning_rate": 0.0001,
+      "loss": 3.0362,
+      "loss/crossentropy": 2.2525784373283386,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25623630434274675,
+      "loss/reg": 0.0,
+      "step": 34520
+    },
+    {
+      "epoch": 0.22717105263157894,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 2.48424072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.097,
+      "loss/crossentropy": 2.3944430470466616,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23495300412178038,
+      "loss/reg": 0.0,
+      "step": 34530
+    },
+    {
+      "epoch": 0.22723684210526315,
+      "grad_norm": 2.75,
+      "grad_norm_var": 2.325169881184896,
+      "learning_rate": 0.0001,
+      "loss": 3.0441,
+      "loss/crossentropy": 2.1533145189285277,
+      "loss/hidden": 2.7984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20518074482679366,
+      "loss/reg": 0.0,
+      "step": 34540
+    },
+    {
+      "epoch": 0.22730263157894737,
+      "grad_norm": 3.453125,
+      "grad_norm_var": 0.6099609375,
+      "learning_rate": 0.0001,
+      "loss": 3.1133,
+      "loss/crossentropy": 2.298123669624329,
+      "loss/hidden": 2.7140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23540952503681184,
+      "loss/reg": 0.0,
+      "step": 34550
+    },
+    {
+      "epoch": 0.22736842105263158,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.169775390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0075,
+      "loss/crossentropy": 2.333153510093689,
+      "loss/hidden": 2.8203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23416389673948287,
+      "loss/reg": 0.0,
+      "step": 34560
+    },
+    {
+      "epoch": 0.2274342105263158,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.14983317057291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0272,
+      "loss/crossentropy": 2.455455815792084,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26056560724973676,
+      "loss/reg": 0.0,
+      "step": 34570
+    },
+    {
+      "epoch": 0.2275,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.17576065063476562,
+      "learning_rate": 0.0001,
+      "loss": 2.9814,
+      "loss/crossentropy": 1.9633507668972014,
+      "loss/hidden": 2.5953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19498348981142044,
+      "loss/reg": 0.0,
+      "step": 34580
+    },
+    {
+      "epoch": 0.22756578947368422,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.1496246337890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0304,
+      "loss/crossentropy": 1.8702698707580567,
+      "loss/hidden": 2.6984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2022824764251709,
+      "loss/reg": 0.0,
+      "step": 34590
+    },
+    {
+      "epoch": 0.22763157894736843,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.1550445556640625,
+      "learning_rate": 0.0001,
+      "loss": 3.027,
+      "loss/crossentropy": 2.588277578353882,
+      "loss/hidden": 2.94375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28010246604681016,
+      "loss/reg": 0.0,
+      "step": 34600
+    },
+    {
+      "epoch": 0.22769736842105262,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.09648335774739583,
+      "learning_rate": 0.0001,
+      "loss": 3.0036,
+      "loss/crossentropy": 2.3437662720680237,
+      "loss/hidden": 2.7453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21182923167943954,
+      "loss/reg": 0.0,
+      "step": 34610
+    },
+    {
+      "epoch": 0.22776315789473683,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.08466389973958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0695,
+      "loss/crossentropy": 2.3003593921661376,
+      "loss/hidden": 2.7984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2246791511774063,
+      "loss/reg": 0.0,
+      "step": 34620
+    },
+    {
+      "epoch": 0.22782894736842105,
+      "grad_norm": 2.625,
+      "grad_norm_var": 2.332779947916667,
+      "learning_rate": 0.0001,
+      "loss": 2.9537,
+      "loss/crossentropy": 2.2143739581108095,
+      "loss/hidden": 2.6265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2086007311940193,
+      "loss/reg": 0.0,
+      "step": 34630
+    },
+    {
+      "epoch": 0.22789473684210526,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 2.5328409830729166,
+      "learning_rate": 0.0001,
+      "loss": 3.0153,
+      "loss/crossentropy": 2.1308518052101135,
+      "loss/hidden": 2.775,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2038565307855606,
+      "loss/reg": 0.0,
+      "step": 34640
+    },
+    {
+      "epoch": 0.22796052631578947,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.36980361938476564,
+      "learning_rate": 0.0001,
+      "loss": 2.9625,
+      "loss/crossentropy": 2.088499677181244,
+      "loss/hidden": 2.8984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.268018501996994,
+      "loss/reg": 0.0,
+      "step": 34650
+    },
+    {
+      "epoch": 0.2280263157894737,
+      "grad_norm": 1962934272.0,
+      "grad_norm_var": 2.408194341792645e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0604,
+      "loss/crossentropy": 2.413490664958954,
+      "loss/hidden": 2.634375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21956574469804763,
+      "loss/reg": 0.0,
+      "step": 34660
+    },
+    {
+      "epoch": 0.2280921052631579,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 2.4081943414143712e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0371,
+      "loss/crossentropy": 2.1357826590538025,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19804717004299163,
+      "loss/reg": 0.0,
+      "step": 34670
+    },
+    {
+      "epoch": 0.22815789473684212,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.16970113118489583,
+      "learning_rate": 0.0001,
+      "loss": 2.9031,
+      "loss/crossentropy": 2.4601009845733643,
+      "loss/hidden": 2.6828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23078858107328415,
+      "loss/reg": 0.0,
+      "step": 34680
+    },
+    {
+      "epoch": 0.22822368421052633,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.0817535400390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0007,
+      "loss/crossentropy": 2.170434999465942,
+      "loss/hidden": 2.7,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23267455995082856,
+      "loss/reg": 0.0,
+      "step": 34690
+    },
+    {
+      "epoch": 0.22828947368421051,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.08743489583333333,
+      "learning_rate": 0.0001,
+      "loss": 2.9957,
+      "loss/crossentropy": 2.378873956203461,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25174411833286287,
+      "loss/reg": 0.0,
+      "step": 34700
+    },
+    {
+      "epoch": 0.22835526315789473,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.0942400614420573,
+      "learning_rate": 0.0001,
+      "loss": 2.9195,
+      "loss/crossentropy": 2.3702409982681276,
+      "loss/hidden": 2.684375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2114430248737335,
+      "loss/reg": 0.0,
+      "step": 34710
+    },
+    {
+      "epoch": 0.22842105263157894,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.1205230712890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9774,
+      "loss/crossentropy": 2.1158406376838683,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2316122904419899,
+      "loss/reg": 0.0,
+      "step": 34720
+    },
+    {
+      "epoch": 0.22848684210526315,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.04747899373372396,
+      "learning_rate": 0.0001,
+      "loss": 2.9868,
+      "loss/crossentropy": 2.2178874969482423,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22333518415689468,
+      "loss/reg": 0.0,
+      "step": 34730
+    },
+    {
+      "epoch": 0.22855263157894737,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.08137105305989584,
+      "learning_rate": 0.0001,
+      "loss": 3.0257,
+      "loss/crossentropy": 2.364703667163849,
+      "loss/hidden": 2.75625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22582925111055374,
+      "loss/reg": 0.0,
+      "step": 34740
+    },
+    {
+      "epoch": 0.22861842105263158,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.0731842041015625,
+      "learning_rate": 0.0001,
+      "loss": 2.9844,
+      "loss/crossentropy": 2.06840842962265,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24444130361080169,
+      "loss/reg": 0.0,
+      "step": 34750
+    },
+    {
+      "epoch": 0.2286842105263158,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.06175308227539063,
+      "learning_rate": 0.0001,
+      "loss": 3.0122,
+      "loss/crossentropy": 2.5480314254760743,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3109170734882355,
+      "loss/reg": 0.0,
+      "step": 34760
+    },
+    {
+      "epoch": 0.22875,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.06448567708333333,
+      "learning_rate": 0.0001,
+      "loss": 2.9767,
+      "loss/crossentropy": 2.4194095849990847,
+      "loss/hidden": 2.709375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23701667934656143,
+      "loss/reg": 0.0,
+      "step": 34770
+    },
+    {
+      "epoch": 0.22881578947368422,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.13319676717122395,
+      "learning_rate": 0.0001,
+      "loss": 3.0117,
+      "loss/crossentropy": 2.330536985397339,
+      "loss/hidden": 2.665625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1930417075753212,
+      "loss/reg": 0.0,
+      "step": 34780
+    },
+    {
+      "epoch": 0.2288815789473684,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.13925755818684896,
+      "learning_rate": 0.0001,
+      "loss": 3.0049,
+      "loss/crossentropy": 2.055137485265732,
+      "loss/hidden": 2.7359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1990390993654728,
+      "loss/reg": 0.0,
+      "step": 34790
+    },
+    {
+      "epoch": 0.22894736842105262,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.0531402587890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9451,
+      "loss/crossentropy": 2.3400762557983397,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2514717549085617,
+      "loss/reg": 0.0,
+      "step": 34800
+    },
+    {
+      "epoch": 0.22901315789473684,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.14586588541666667,
+      "learning_rate": 0.0001,
+      "loss": 3.0989,
+      "loss/crossentropy": 2.3216898441314697,
+      "loss/hidden": 2.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20793744474649428,
+      "loss/reg": 0.0,
+      "step": 34810
+    },
+    {
+      "epoch": 0.22907894736842105,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.23554280598958333,
+      "learning_rate": 0.0001,
+      "loss": 3.1053,
+      "loss/crossentropy": 2.368983769416809,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23178687542676926,
+      "loss/reg": 0.0,
+      "step": 34820
+    },
+    {
+      "epoch": 0.22914473684210526,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.09830322265625,
+      "learning_rate": 0.0001,
+      "loss": 2.9238,
+      "loss/crossentropy": 2.1155460715293883,
+      "loss/hidden": 2.659375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21411084979772568,
+      "loss/reg": 0.0,
+      "step": 34830
+    },
+    {
+      "epoch": 0.22921052631578948,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.6712799072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0934,
+      "loss/crossentropy": 1.9747644543647767,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22649588584899902,
+      "loss/reg": 0.0,
+      "step": 34840
+    },
+    {
+      "epoch": 0.2292763157894737,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.6342081705729167,
+      "learning_rate": 0.0001,
+      "loss": 2.968,
+      "loss/crossentropy": 2.282280433177948,
+      "loss/hidden": 2.8265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24287160784006118,
+      "loss/reg": 0.0,
+      "step": 34850
+    },
+    {
+      "epoch": 0.2293421052631579,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.676806640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0144,
+      "loss/crossentropy": 2.423893141746521,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27746856659650804,
+      "loss/reg": 0.0,
+      "step": 34860
+    },
+    {
+      "epoch": 0.22940789473684212,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.15266927083333334,
+      "learning_rate": 0.0001,
+      "loss": 2.9868,
+      "loss/crossentropy": 2.2340946078300474,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23156266957521437,
+      "loss/reg": 0.0,
+      "step": 34870
+    },
+    {
+      "epoch": 0.2294736842105263,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.12720947265625,
+      "learning_rate": 0.0001,
+      "loss": 2.9166,
+      "loss/crossentropy": 2.3554929852485658,
+      "loss/hidden": 2.70625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2240227773785591,
+      "loss/reg": 0.0,
+      "step": 34880
+    },
+    {
+      "epoch": 0.22953947368421052,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.1996978759765625,
+      "learning_rate": 0.0001,
+      "loss": 2.9409,
+      "loss/crossentropy": 2.1706709206104278,
+      "loss/hidden": 2.6015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1888253793120384,
+      "loss/reg": 0.0,
+      "step": 34890
+    },
+    {
+      "epoch": 0.22960526315789473,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.26612955729166665,
+      "learning_rate": 0.0001,
+      "loss": 3.0791,
+      "loss/crossentropy": 2.2094788432121275,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21013466268777847,
+      "loss/reg": 0.0,
+      "step": 34900
+    },
+    {
+      "epoch": 0.22967105263157894,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.3094716389973958,
+      "learning_rate": 0.0001,
+      "loss": 3.0647,
+      "loss/crossentropy": 2.008031153678894,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2073305867612362,
+      "loss/reg": 0.0,
+      "step": 34910
+    },
+    {
+      "epoch": 0.22973684210526316,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.16913960774739584,
+      "learning_rate": 0.0001,
+      "loss": 2.9933,
+      "loss/crossentropy": 2.3257227003574372,
+      "loss/hidden": 2.7359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2534154921770096,
+      "loss/reg": 0.0,
+      "step": 34920
+    },
+    {
+      "epoch": 0.22980263157894737,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.09401041666666667,
+      "learning_rate": 0.0001,
+      "loss": 3.078,
+      "loss/crossentropy": 2.299355685710907,
+      "loss/hidden": 2.890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4137351721525192,
+      "loss/reg": 0.0,
+      "step": 34930
+    },
+    {
+      "epoch": 0.22986842105263158,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.059163411458333336,
+      "learning_rate": 0.0001,
+      "loss": 3.066,
+      "loss/crossentropy": 2.063184082508087,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2067345403134823,
+      "loss/reg": 0.0,
+      "step": 34940
+    },
+    {
+      "epoch": 0.2299342105263158,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.0319000244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0384,
+      "loss/crossentropy": 2.3379054069519043,
+      "loss/hidden": 2.625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21520548760890962,
+      "loss/reg": 0.0,
+      "step": 34950
+    },
+    {
+      "epoch": 0.23,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.04029541015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0224,
+      "loss/crossentropy": 2.2715999722480773,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23961079716682435,
+      "loss/reg": 0.0,
+      "step": 34960
+    },
+    {
+      "epoch": 0.23006578947368422,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.017545572916666665,
+      "learning_rate": 0.0001,
+      "loss": 2.9942,
+      "loss/crossentropy": 2.547387886047363,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2642354011535645,
+      "loss/reg": 0.0,
+      "step": 34970
+    },
+    {
+      "epoch": 0.2301315789473684,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.0195465087890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9968,
+      "loss/crossentropy": 2.3052905440330504,
+      "loss/hidden": 2.6859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2478991135954857,
+      "loss/reg": 0.0,
+      "step": 34980
+    },
+    {
+      "epoch": 0.23019736842105262,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.03509114583333333,
+      "learning_rate": 0.0001,
+      "loss": 2.9466,
+      "loss/crossentropy": 2.136173462867737,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23240896165370942,
+      "loss/reg": 0.0,
+      "step": 34990
+    },
+    {
+      "epoch": 0.23026315789473684,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.04448140462239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0289,
+      "loss/crossentropy": 2.2616278886795045,
+      "loss/hidden": 2.6890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2103984072804451,
+      "loss/reg": 0.0,
+      "step": 35000
+    },
+    {
+      "epoch": 0.23032894736842105,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.055464680989583334,
+      "learning_rate": 0.0001,
+      "loss": 3.0574,
+      "loss/crossentropy": 2.234957480430603,
+      "loss/hidden": 2.765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22579004764556884,
+      "loss/reg": 0.0,
+      "step": 35010
+    },
+    {
+      "epoch": 0.23039473684210526,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.0838531494140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0444,
+      "loss/crossentropy": 2.2260616183280946,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2393326446413994,
+      "loss/reg": 0.0,
+      "step": 35020
+    },
+    {
+      "epoch": 0.23046052631578948,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.110595703125,
+      "learning_rate": 0.0001,
+      "loss": 3.0705,
+      "loss/crossentropy": 2.2648324608802795,
+      "loss/hidden": 2.8859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2452955961227417,
+      "loss/reg": 0.0,
+      "step": 35030
+    },
+    {
+      "epoch": 0.2305263157894737,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.05959879557291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0123,
+      "loss/crossentropy": 2.3035391211509704,
+      "loss/hidden": 2.675,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20398545116186143,
+      "loss/reg": 0.0,
+      "step": 35040
+    },
+    {
+      "epoch": 0.2305921052631579,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.10580952962239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0008,
+      "loss/crossentropy": 2.467655122280121,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2141210839152336,
+      "loss/reg": 0.0,
+      "step": 35050
+    },
+    {
+      "epoch": 0.23065789473684212,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.12373046875,
+      "learning_rate": 0.0001,
+      "loss": 2.9259,
+      "loss/crossentropy": 2.4413907408714293,
+      "loss/hidden": 2.646875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21587167531251908,
+      "loss/reg": 0.0,
+      "step": 35060
+    },
+    {
+      "epoch": 0.2307236842105263,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.18179931640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0733,
+      "loss/crossentropy": 2.5797677636146545,
+      "loss/hidden": 3.096875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3117774799466133,
+      "loss/reg": 0.0,
+      "step": 35070
+    },
+    {
+      "epoch": 0.23078947368421052,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 8.836572265625,
+      "learning_rate": 0.0001,
+      "loss": 2.98,
+      "loss/crossentropy": 2.0451542496681214,
+      "loss/hidden": 2.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25981259644031524,
+      "loss/reg": 0.0,
+      "step": 35080
+    },
+    {
+      "epoch": 0.23085526315789473,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 8.937858072916667,
+      "learning_rate": 0.0001,
+      "loss": 2.9729,
+      "loss/crossentropy": 2.4952213764190674,
+      "loss/hidden": 2.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27273062616586685,
+      "loss/reg": 0.0,
+      "step": 35090
+    },
+    {
+      "epoch": 0.23092105263157894,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.039632161458333336,
+      "learning_rate": 0.0001,
+      "loss": 3.0086,
+      "loss/crossentropy": 2.1235549569129946,
+      "loss/hidden": 2.846875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2294046923518181,
+      "loss/reg": 0.0,
+      "step": 35100
+    },
+    {
+      "epoch": 0.23098684210526316,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.2903065999348958,
+      "learning_rate": 0.0001,
+      "loss": 3.004,
+      "loss/crossentropy": 2.2860934376716613,
+      "loss/hidden": 3.009375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23025134950876236,
+      "loss/reg": 0.0,
+      "step": 35110
+    },
+    {
+      "epoch": 0.23105263157894737,
+      "grad_norm": 3.125,
+      "grad_norm_var": 0.1121978759765625,
+      "learning_rate": 0.0001,
+      "loss": 2.9751,
+      "loss/crossentropy": 2.2237872958183287,
+      "loss/hidden": 2.8,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2481079339981079,
+      "loss/reg": 0.0,
+      "step": 35120
+    },
+    {
+      "epoch": 0.23111842105263158,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.12190348307291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0254,
+      "loss/crossentropy": 2.367337203025818,
+      "loss/hidden": 2.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24131873100996018,
+      "loss/reg": 0.0,
+      "step": 35130
+    },
+    {
+      "epoch": 0.2311842105263158,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.061310831705729166,
+      "learning_rate": 0.0001,
+      "loss": 3.0044,
+      "loss/crossentropy": 2.287555253505707,
+      "loss/hidden": 2.7078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24582310616970063,
+      "loss/reg": 0.0,
+      "step": 35140
+    },
+    {
+      "epoch": 0.23125,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.0275543212890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0166,
+      "loss/crossentropy": 2.3102270364761353,
+      "loss/hidden": 2.734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22130452394485473,
+      "loss/reg": 0.0,
+      "step": 35150
+    },
+    {
+      "epoch": 0.2313157894736842,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.022557576497395832,
+      "learning_rate": 0.0001,
+      "loss": 3.0183,
+      "loss/crossentropy": 2.3422921657562257,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26338334679603576,
+      "loss/reg": 0.0,
+      "step": 35160
+    },
+    {
+      "epoch": 0.2313815789473684,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.0848297119140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0288,
+      "loss/crossentropy": 2.1819352507591248,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21551729440689088,
+      "loss/reg": 0.0,
+      "step": 35170
+    },
+    {
+      "epoch": 0.23144736842105262,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.17290751139322916,
+      "learning_rate": 0.0001,
+      "loss": 3.0282,
+      "loss/crossentropy": 2.512664186954498,
+      "loss/hidden": 2.65625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20309768319129945,
+      "loss/reg": 0.0,
+      "step": 35180
+    },
+    {
+      "epoch": 0.23151315789473684,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 3.398986258287209e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1551,
+      "loss/crossentropy": 2.3065868735313417,
+      "loss/hidden": 2.790625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2547546997666359,
+      "loss/reg": 0.0,
+      "step": 35190
+    },
+    {
+      "epoch": 0.23157894736842105,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.2131744384765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0611,
+      "loss/crossentropy": 2.5265056133270263,
+      "loss/hidden": 2.7375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21266352534294128,
+      "loss/reg": 0.0,
+      "step": 35200
+    },
+    {
+      "epoch": 0.23164473684210526,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.4802480061848958,
+      "learning_rate": 0.0001,
+      "loss": 3.1018,
+      "loss/crossentropy": 2.378112018108368,
+      "loss/hidden": 2.68125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2218308851122856,
+      "loss/reg": 0.0,
+      "step": 35210
+    },
+    {
+      "epoch": 0.23171052631578948,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.28329671223958336,
+      "learning_rate": 0.0001,
+      "loss": 2.9809,
+      "loss/crossentropy": 2.3841970801353454,
+      "loss/hidden": 2.646875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20064097195863723,
+      "loss/reg": 0.0,
+      "step": 35220
+    },
+    {
+      "epoch": 0.2317763157894737,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.13202718098958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0415,
+      "loss/crossentropy": 2.2034762263298036,
+      "loss/hidden": 2.8,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26952334195375444,
+      "loss/reg": 0.0,
+      "step": 35230
+    },
+    {
+      "epoch": 0.2318421052631579,
+      "grad_norm": 3.484375,
+      "grad_norm_var": 0.17197850545247395,
+      "learning_rate": 0.0001,
+      "loss": 2.9833,
+      "loss/crossentropy": 2.333740735054016,
+      "loss/hidden": 2.6328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2028803788125515,
+      "loss/reg": 0.0,
+      "step": 35240
+    },
+    {
+      "epoch": 0.23190789473684212,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.12504247029622395,
+      "learning_rate": 0.0001,
+      "loss": 3.0428,
+      "loss/crossentropy": 1.9891301035881042,
+      "loss/hidden": 2.871875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.33240934908390046,
+      "loss/reg": 0.0,
+      "step": 35250
+    },
+    {
+      "epoch": 0.2319736842105263,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.0198394775390625,
+      "learning_rate": 0.0001,
+      "loss": 2.9679,
+      "loss/crossentropy": 2.394805371761322,
+      "loss/hidden": 2.65,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2079624727368355,
+      "loss/reg": 0.0,
+      "step": 35260
+    },
+    {
+      "epoch": 0.23203947368421052,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.07119140625,
+      "learning_rate": 0.0001,
+      "loss": 2.9843,
+      "loss/crossentropy": 2.466187059879303,
+      "loss/hidden": 2.665625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21209322810173034,
+      "loss/reg": 0.0,
+      "step": 35270
+    },
+    {
+      "epoch": 0.23210526315789473,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.7116737365722656,
+      "learning_rate": 0.0001,
+      "loss": 2.9743,
+      "loss/crossentropy": 2.2582387804985045,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2430204778909683,
+      "loss/reg": 0.0,
+      "step": 35280
+    },
+    {
+      "epoch": 0.23217105263157894,
+      "grad_norm": 3.28125,
+      "grad_norm_var": 3.206175898909409e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1474,
+      "loss/crossentropy": 2.2272575974464415,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20790190547704696,
+      "loss/reg": 0.0,
+      "step": 35290
+    },
+    {
+      "epoch": 0.23223684210526316,
+      "grad_norm": 2.75,
+      "grad_norm_var": 3.206175899148288e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0553,
+      "loss/crossentropy": 2.3405486226081846,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24497875720262527,
+      "loss/reg": 0.0,
+      "step": 35300
+    },
+    {
+      "epoch": 0.23230263157894737,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.12532145182291668,
+      "learning_rate": 0.0001,
+      "loss": 3.0759,
+      "loss/crossentropy": 2.316471701860428,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23606702983379363,
+      "loss/reg": 0.0,
+      "step": 35310
+    },
+    {
+      "epoch": 0.23236842105263159,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.07993876139322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0771,
+      "loss/crossentropy": 2.068638467788696,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2723642893135548,
+      "loss/reg": 0.0,
+      "step": 35320
+    },
+    {
+      "epoch": 0.2324342105263158,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.1643463134765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0607,
+      "loss/crossentropy": 2.209253963828087,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2534912425093353,
+      "loss/reg": 0.0,
+      "step": 35330
+    },
+    {
+      "epoch": 0.2325,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.0254302978515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0043,
+      "loss/crossentropy": 2.3624029099941253,
+      "loss/hidden": 2.7421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25295102000236513,
+      "loss/reg": 0.0,
+      "step": 35340
+    },
+    {
+      "epoch": 0.2325657894736842,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.04498291015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0386,
+      "loss/crossentropy": 2.3273918867111205,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22172463536262513,
+      "loss/reg": 0.0,
+      "step": 35350
+    },
+    {
+      "epoch": 0.2326315789473684,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.055403645833333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0584,
+      "loss/crossentropy": 1.895402479171753,
+      "loss/hidden": 3.0015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23743433207273484,
+      "loss/reg": 0.0,
+      "step": 35360
+    },
+    {
+      "epoch": 0.23269736842105262,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.056962076822916666,
+      "learning_rate": 0.0001,
+      "loss": 3.1011,
+      "loss/crossentropy": 2.0022137641906737,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21825481653213502,
+      "loss/reg": 0.0,
+      "step": 35370
+    },
+    {
+      "epoch": 0.23276315789473684,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.08046468098958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0227,
+      "loss/crossentropy": 1.9984350323677063,
+      "loss/hidden": 2.9140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2356773540377617,
+      "loss/reg": 0.0,
+      "step": 35380
+    },
+    {
+      "epoch": 0.23282894736842105,
+      "grad_norm": 3.140625,
+      "grad_norm_var": 0.13664957682291667,
+      "learning_rate": 0.0001,
+      "loss": 3.0069,
+      "loss/crossentropy": 2.1517210602760315,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22341941446065902,
+      "loss/reg": 0.0,
+      "step": 35390
+    },
+    {
+      "epoch": 0.23289473684210527,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.26809794108072915,
+      "learning_rate": 0.0001,
+      "loss": 3.0346,
+      "loss/crossentropy": 2.380366015434265,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2661885678768158,
+      "loss/reg": 0.0,
+      "step": 35400
+    },
+    {
+      "epoch": 0.23296052631578948,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.036149088541666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0274,
+      "loss/crossentropy": 2.084783911705017,
+      "loss/hidden": 2.765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22119598612189292,
+      "loss/reg": 0.0,
+      "step": 35410
+    },
+    {
+      "epoch": 0.2330263157894737,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.21288960774739582,
+      "learning_rate": 0.0001,
+      "loss": 2.9574,
+      "loss/crossentropy": 2.288650333881378,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2464343011379242,
+      "loss/reg": 0.0,
+      "step": 35420
+    },
+    {
+      "epoch": 0.2330921052631579,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.22062174479166666,
+      "learning_rate": 0.0001,
+      "loss": 3.0144,
+      "loss/crossentropy": 2.1648478746414184,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2272023007273674,
+      "loss/reg": 0.0,
+      "step": 35430
+    },
+    {
+      "epoch": 0.2331578947368421,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.07825113932291666,
+      "learning_rate": 0.0001,
+      "loss": 2.9664,
+      "loss/crossentropy": 2.4016910076141356,
+      "loss/hidden": 2.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19540313482284546,
+      "loss/reg": 0.0,
+      "step": 35440
+    },
+    {
+      "epoch": 0.2332236842105263,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.12473526000976562,
+      "learning_rate": 0.0001,
+      "loss": 3.0054,
+      "loss/crossentropy": 2.2149259626865385,
+      "loss/hidden": 2.6640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.211626535654068,
+      "loss/reg": 0.0,
+      "step": 35450
+    },
+    {
+      "epoch": 0.23328947368421052,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.2657244364420573,
+      "learning_rate": 0.0001,
+      "loss": 3.022,
+      "loss/crossentropy": 2.0099823474884033,
+      "loss/hidden": 2.8078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21212436258792877,
+      "loss/reg": 0.0,
+      "step": 35460
+    },
+    {
+      "epoch": 0.23335526315789473,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.035560862223307295,
+      "learning_rate": 0.0001,
+      "loss": 2.959,
+      "loss/crossentropy": 2.149253064393997,
+      "loss/hidden": 2.6546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19234723448753357,
+      "loss/reg": 0.0,
+      "step": 35470
+    },
+    {
+      "epoch": 0.23342105263157895,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.196630859375,
+      "learning_rate": 0.0001,
+      "loss": 3.0157,
+      "loss/crossentropy": 2.3657627940177917,
+      "loss/hidden": 2.753125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.223566497862339,
+      "loss/reg": 0.0,
+      "step": 35480
+    },
+    {
+      "epoch": 0.23348684210526316,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.024527994791666667,
+      "learning_rate": 0.0001,
+      "loss": 2.9839,
+      "loss/crossentropy": 2.4007309794425966,
+      "loss/hidden": 2.596875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19649036675691606,
+      "loss/reg": 0.0,
+      "step": 35490
+    },
+    {
+      "epoch": 0.23355263157894737,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.05445556640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0323,
+      "loss/crossentropy": 2.414763641357422,
+      "loss/hidden": 2.9296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2565377399325371,
+      "loss/reg": 0.0,
+      "step": 35500
+    },
+    {
+      "epoch": 0.2336184210526316,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.04409077962239583,
+      "learning_rate": 0.0001,
+      "loss": 2.9834,
+      "loss/crossentropy": 2.272168481349945,
+      "loss/hidden": 2.6125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2014573760330677,
+      "loss/reg": 0.0,
+      "step": 35510
+    },
+    {
+      "epoch": 0.2336842105263158,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.091162109375,
+      "learning_rate": 0.0001,
+      "loss": 3.0118,
+      "loss/crossentropy": 2.4486998319625854,
+      "loss/hidden": 2.7515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2566034272313118,
+      "loss/reg": 0.0,
+      "step": 35520
+    },
+    {
+      "epoch": 0.23375,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.11804097493489583,
+      "learning_rate": 0.0001,
+      "loss": 2.9697,
+      "loss/crossentropy": 2.2795116662979127,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20817176550626754,
+      "loss/reg": 0.0,
+      "step": 35530
+    },
+    {
+      "epoch": 0.2338157894736842,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.05882059733072917,
+      "learning_rate": 0.0001,
+      "loss": 2.9864,
+      "loss/crossentropy": 2.1612624883651734,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.234655499458313,
+      "loss/reg": 0.0,
+      "step": 35540
+    },
+    {
+      "epoch": 0.2338815789473684,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.051545206705729166,
+      "learning_rate": 0.0001,
+      "loss": 2.9369,
+      "loss/crossentropy": 2.342636692523956,
+      "loss/hidden": 2.6953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2195874720811844,
+      "loss/reg": 0.0,
+      "step": 35550
+    },
+    {
+      "epoch": 0.23394736842105263,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.2511678059895833,
+      "learning_rate": 0.0001,
+      "loss": 2.9652,
+      "loss/crossentropy": 2.4172530651092528,
+      "loss/hidden": 2.675,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22586058229207992,
+      "loss/reg": 0.0,
+      "step": 35560
+    },
+    {
+      "epoch": 0.23401315789473684,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.21130269368489582,
+      "learning_rate": 0.0001,
+      "loss": 2.9721,
+      "loss/crossentropy": 2.19435909986496,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22134042978286744,
+      "loss/reg": 0.0,
+      "step": 35570
+    },
+    {
+      "epoch": 0.23407894736842105,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.12469075520833334,
+      "learning_rate": 0.0001,
+      "loss": 3.0487,
+      "loss/crossentropy": 2.491572880744934,
+      "loss/hidden": 2.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2359505832195282,
+      "loss/reg": 0.0,
+      "step": 35580
+    },
+    {
+      "epoch": 0.23414473684210527,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.24637832641601562,
+      "learning_rate": 0.0001,
+      "loss": 3.0188,
+      "loss/crossentropy": 2.0867708206176756,
+      "loss/hidden": 2.946875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29949132055044175,
+      "loss/reg": 0.0,
+      "step": 35590
+    },
+    {
+      "epoch": 0.23421052631578948,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.09173075358072917,
+      "learning_rate": 0.0001,
+      "loss": 3.0162,
+      "loss/crossentropy": 2.500631070137024,
+      "loss/hidden": 2.6359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23038013875484467,
+      "loss/reg": 0.0,
+      "step": 35600
+    },
+    {
+      "epoch": 0.2342763157894737,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.09622294108072917,
+      "learning_rate": 0.0001,
+      "loss": 2.9738,
+      "loss/crossentropy": 2.330187511444092,
+      "loss/hidden": 2.6375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21412627547979354,
+      "loss/reg": 0.0,
+      "step": 35610
+    },
+    {
+      "epoch": 0.2343421052631579,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.06207275390625,
+      "learning_rate": 0.0001,
+      "loss": 2.9646,
+      "loss/crossentropy": 2.0748894095420836,
+      "loss/hidden": 2.9234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2262689083814621,
+      "loss/reg": 0.0,
+      "step": 35620
+    },
+    {
+      "epoch": 0.2344078947368421,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.0324859619140625,
+      "learning_rate": 0.0001,
+      "loss": 2.9761,
+      "loss/crossentropy": 2.415894079208374,
+      "loss/hidden": 2.6671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20092394277453424,
+      "loss/reg": 0.0,
+      "step": 35630
+    },
+    {
+      "epoch": 0.2344736842105263,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.1505859375,
+      "learning_rate": 0.0001,
+      "loss": 3.0498,
+      "loss/crossentropy": 2.2875362753868105,
+      "loss/hidden": 2.884375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24587944149971008,
+      "loss/reg": 0.0,
+      "step": 35640
+    },
+    {
+      "epoch": 0.23453947368421052,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.16611226399739584,
+      "learning_rate": 0.0001,
+      "loss": 2.9052,
+      "loss/crossentropy": 2.4340951919555662,
+      "loss/hidden": 2.6359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22377320230007172,
+      "loss/reg": 0.0,
+      "step": 35650
+    },
+    {
+      "epoch": 0.23460526315789473,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.06776936848958333,
+      "learning_rate": 0.0001,
+      "loss": 2.9088,
+      "loss/crossentropy": 2.3177594542503357,
+      "loss/hidden": 2.6875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2174980953335762,
+      "loss/reg": 0.0,
+      "step": 35660
+    },
+    {
+      "epoch": 0.23467105263157895,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.0744049072265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0171,
+      "loss/crossentropy": 2.2726083517074587,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30478386878967284,
+      "loss/reg": 0.0,
+      "step": 35670
+    },
+    {
+      "epoch": 0.23473684210526316,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.0516754150390625,
+      "learning_rate": 0.0001,
+      "loss": 3.072,
+      "loss/crossentropy": 2.2021562099456786,
+      "loss/hidden": 3.0625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27211851328611375,
+      "loss/reg": 0.0,
+      "step": 35680
+    },
+    {
+      "epoch": 0.23480263157894737,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.05318603515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0061,
+      "loss/crossentropy": 2.2337870597839355,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23637954592704774,
+      "loss/reg": 0.0,
+      "step": 35690
+    },
+    {
+      "epoch": 0.2348684210526316,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.21002197265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0582,
+      "loss/crossentropy": 1.9850136280059814,
+      "loss/hidden": 3.0078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2096769317984581,
+      "loss/reg": 0.0,
+      "step": 35700
+    },
+    {
+      "epoch": 0.2349342105263158,
+      "grad_norm": 6.125,
+      "grad_norm_var": 1.515673828125,
+      "learning_rate": 0.0001,
+      "loss": 2.9576,
+      "loss/crossentropy": 2.496402633190155,
+      "loss/hidden": 2.690625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22533617317676544,
+      "loss/reg": 0.0,
+      "step": 35710
+    },
+    {
+      "epoch": 0.235,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.95963134765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0797,
+      "loss/crossentropy": 2.140752410888672,
+      "loss/hidden": 2.99375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20643131881952287,
+      "loss/reg": 0.0,
+      "step": 35720
+    },
+    {
+      "epoch": 0.2350657894736842,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.19202067057291666,
+      "learning_rate": 0.0001,
+      "loss": 2.9408,
+      "loss/crossentropy": 1.9449581146240233,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1967112362384796,
+      "loss/reg": 0.0,
+      "step": 35730
+    },
+    {
+      "epoch": 0.2351315789473684,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.718365224202474,
+      "learning_rate": 0.0001,
+      "loss": 2.873,
+      "loss/crossentropy": 2.059830403327942,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1984706148505211,
+      "loss/reg": 0.0,
+      "step": 35740
+    },
+    {
+      "epoch": 0.23519736842105263,
+      "grad_norm": 1.90625,
+      "grad_norm_var": 0.7783404032389323,
+      "learning_rate": 0.0001,
+      "loss": 2.9617,
+      "loss/crossentropy": 2.5269897818565368,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24567708522081375,
+      "loss/reg": 0.0,
+      "step": 35750
+    },
+    {
+      "epoch": 0.23526315789473684,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.3910441080729167,
+      "learning_rate": 0.0001,
+      "loss": 3.049,
+      "loss/crossentropy": 2.2286418437957765,
+      "loss/hidden": 3.084375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22506199181079864,
+      "loss/reg": 0.0,
+      "step": 35760
+    },
+    {
+      "epoch": 0.23532894736842105,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.20345052083333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0262,
+      "loss/crossentropy": 2.321590745449066,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22042199671268464,
+      "loss/reg": 0.0,
+      "step": 35770
+    },
+    {
+      "epoch": 0.23539473684210527,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.6254981994628906,
+      "learning_rate": 0.0001,
+      "loss": 2.9981,
+      "loss/crossentropy": 2.3971486926078795,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21692513525485993,
+      "loss/reg": 0.0,
+      "step": 35780
+    },
+    {
+      "epoch": 0.23546052631578948,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.2902809143066406,
+      "learning_rate": 0.0001,
+      "loss": 2.931,
+      "loss/crossentropy": 2.3023766756057737,
+      "loss/hidden": 2.7078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2162985995411873,
+      "loss/reg": 0.0,
+      "step": 35790
+    },
+    {
+      "epoch": 0.2355263157894737,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.16299540201822918,
+      "learning_rate": 0.0001,
+      "loss": 3.0115,
+      "loss/crossentropy": 2.355253207683563,
+      "loss/hidden": 2.7515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24053574204444886,
+      "loss/reg": 0.0,
+      "step": 35800
+    },
+    {
+      "epoch": 0.2355921052631579,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.10937398274739583,
+      "learning_rate": 0.0001,
+      "loss": 2.9729,
+      "loss/crossentropy": 2.493846869468689,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2182188354432583,
+      "loss/reg": 0.0,
+      "step": 35810
+    },
+    {
+      "epoch": 0.2356578947368421,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.46917317708333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0177,
+      "loss/crossentropy": 2.3108774185180665,
+      "loss/hidden": 2.796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.17969555854797364,
+      "loss/reg": 0.0,
+      "step": 35820
+    },
+    {
+      "epoch": 0.2357236842105263,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.29129231770833336,
+      "learning_rate": 0.0001,
+      "loss": 3.0199,
+      "loss/crossentropy": 2.1752161145210267,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25736078470945356,
+      "loss/reg": 0.0,
+      "step": 35830
+    },
+    {
+      "epoch": 0.23578947368421052,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.15127665201822918,
+      "learning_rate": 0.0001,
+      "loss": 3.0057,
+      "loss/crossentropy": 2.21168737411499,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24892309606075286,
+      "loss/reg": 0.0,
+      "step": 35840
+    },
+    {
+      "epoch": 0.23585526315789473,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.38827718098958336,
+      "learning_rate": 0.0001,
+      "loss": 2.9855,
+      "loss/crossentropy": 2.4601115822792052,
+      "loss/hidden": 2.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23503982722759248,
+      "loss/reg": 0.0,
+      "step": 35850
+    },
+    {
+      "epoch": 0.23592105263157895,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.0817779541015625,
+      "learning_rate": 0.0001,
+      "loss": 2.9943,
+      "loss/crossentropy": 2.4704690098762514,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2563426047563553,
+      "loss/reg": 0.0,
+      "step": 35860
+    },
+    {
+      "epoch": 0.23598684210526316,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.16731541951497395,
+      "learning_rate": 0.0001,
+      "loss": 3.0406,
+      "loss/crossentropy": 2.6593142986297607,
+      "loss/hidden": 3.028125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2891248628497124,
+      "loss/reg": 0.0,
+      "step": 35870
+    },
+    {
+      "epoch": 0.23605263157894738,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.13696263631184896,
+      "learning_rate": 0.0001,
+      "loss": 3.0187,
+      "loss/crossentropy": 2.4739042282104493,
+      "loss/hidden": 2.7046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23506251722574234,
+      "loss/reg": 0.0,
+      "step": 35880
+    },
+    {
+      "epoch": 0.2361184210526316,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.159375,
+      "learning_rate": 0.0001,
+      "loss": 3.0207,
+      "loss/crossentropy": 2.2195907831192017,
+      "loss/hidden": 2.6703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20974744409322738,
+      "loss/reg": 0.0,
+      "step": 35890
+    },
+    {
+      "epoch": 0.2361842105263158,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.2710039774576823,
+      "learning_rate": 0.0001,
+      "loss": 2.9166,
+      "loss/crossentropy": 2.1691187381744386,
+      "loss/hidden": 2.7046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20902955271303653,
+      "loss/reg": 0.0,
+      "step": 35900
+    },
+    {
+      "epoch": 0.23625,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.021848297119140624,
+      "learning_rate": 0.0001,
+      "loss": 2.9421,
+      "loss/crossentropy": 2.380651593208313,
+      "loss/hidden": 2.64375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21168311238288878,
+      "loss/reg": 0.0,
+      "step": 35910
+    },
+    {
+      "epoch": 0.2363157894736842,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.0631256103515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9784,
+      "loss/crossentropy": 2.194112575054169,
+      "loss/hidden": 2.925,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2107112467288971,
+      "loss/reg": 0.0,
+      "step": 35920
+    },
+    {
+      "epoch": 0.23638157894736841,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.19700419108072917,
+      "learning_rate": 0.0001,
+      "loss": 3.0817,
+      "loss/crossentropy": 2.289298951625824,
+      "loss/hidden": 2.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1993821457028389,
+      "loss/reg": 0.0,
+      "step": 35930
+    },
+    {
+      "epoch": 0.23644736842105263,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.07625325520833333,
+      "learning_rate": 0.0001,
+      "loss": 2.9474,
+      "loss/crossentropy": 2.4226235032081602,
+      "loss/hidden": 2.7390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2387748032808304,
+      "loss/reg": 0.0,
+      "step": 35940
+    },
+    {
+      "epoch": 0.23651315789473684,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.04461263020833333,
+      "learning_rate": 0.0001,
+      "loss": 2.9628,
+      "loss/crossentropy": 2.380166971683502,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2143217995762825,
+      "loss/reg": 0.0,
+      "step": 35950
+    },
+    {
+      "epoch": 0.23657894736842106,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.1749908447265625,
+      "learning_rate": 0.0001,
+      "loss": 2.9418,
+      "loss/crossentropy": 2.286650228500366,
+      "loss/hidden": 2.5953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21095365434885024,
+      "loss/reg": 0.0,
+      "step": 35960
+    },
+    {
+      "epoch": 0.23664473684210527,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.11900634765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0537,
+      "loss/crossentropy": 2.2950440287590026,
+      "loss/hidden": 2.8796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2576613754034042,
+      "loss/reg": 0.0,
+      "step": 35970
+    },
+    {
+      "epoch": 0.23671052631578948,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.12654520670572916,
+      "learning_rate": 0.0001,
+      "loss": 2.9324,
+      "loss/crossentropy": 2.236435151100159,
+      "loss/hidden": 2.6484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20467466115951538,
+      "loss/reg": 0.0,
+      "step": 35980
+    },
+    {
+      "epoch": 0.2367763157894737,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.12952473958333333,
+      "learning_rate": 0.0001,
+      "loss": 2.9877,
+      "loss/crossentropy": 2.503733921051025,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21396732181310654,
+      "loss/reg": 0.0,
+      "step": 35990
+    },
+    {
+      "epoch": 0.23684210526315788,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.10692952473958334,
+      "learning_rate": 0.0001,
+      "loss": 2.9584,
+      "loss/crossentropy": 2.493756449222565,
+      "loss/hidden": 2.9109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23922124207019807,
+      "loss/reg": 0.0,
+      "step": 36000
+    },
+    {
+      "epoch": 0.2369078947368421,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.92545166015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0174,
+      "loss/crossentropy": 1.8596992015838623,
+      "loss/hidden": 2.6828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19636918008327484,
+      "loss/reg": 0.0,
+      "step": 36010
+    },
+    {
+      "epoch": 0.2369736842105263,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.9360260009765625,
+      "learning_rate": 0.0001,
+      "loss": 2.9994,
+      "loss/crossentropy": 2.239300674200058,
+      "loss/hidden": 2.75625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21985187232494355,
+      "loss/reg": 0.0,
+      "step": 36020
+    },
+    {
+      "epoch": 0.23703947368421052,
+      "grad_norm": 4.625,
+      "grad_norm_var": 2.2611793518066405,
+      "learning_rate": 0.0001,
+      "loss": 2.987,
+      "loss/crossentropy": 2.22670122385025,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22871969491243363,
+      "loss/reg": 0.0,
+      "step": 36030
+    },
+    {
+      "epoch": 0.23710526315789474,
+      "grad_norm": 3.0,
+      "grad_norm_var": 0.4241167704264323,
+      "learning_rate": 0.0001,
+      "loss": 3.0196,
+      "loss/crossentropy": 2.0241116285324097,
+      "loss/hidden": 2.94375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24767006561160088,
+      "loss/reg": 0.0,
+      "step": 36040
+    },
+    {
+      "epoch": 0.23717105263157895,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.30030924479166665,
+      "learning_rate": 0.0001,
+      "loss": 2.9484,
+      "loss/crossentropy": 2.086419093608856,
+      "loss/hidden": 2.696875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1900019347667694,
+      "loss/reg": 0.0,
+      "step": 36050
+    },
+    {
+      "epoch": 0.23723684210526316,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.1086822509765625,
+      "learning_rate": 0.0001,
+      "loss": 2.9919,
+      "loss/crossentropy": 2.1598266899585723,
+      "loss/hidden": 2.634375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26148284450173376,
+      "loss/reg": 0.0,
+      "step": 36060
+    },
+    {
+      "epoch": 0.23730263157894738,
+      "grad_norm": 3.203125,
+      "grad_norm_var": 0.10439453125,
+      "learning_rate": 0.0001,
+      "loss": 3.0653,
+      "loss/crossentropy": 2.1726396083831787,
+      "loss/hidden": 2.8859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2597889766097069,
+      "loss/reg": 0.0,
+      "step": 36070
+    },
+    {
+      "epoch": 0.2373684210526316,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.10669657389322916,
+      "learning_rate": 0.0001,
+      "loss": 2.9768,
+      "loss/crossentropy": 2.1650060296058653,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22962034344673157,
+      "loss/reg": 0.0,
+      "step": 36080
+    },
+    {
+      "epoch": 0.2374342105263158,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.04478759765625,
+      "learning_rate": 0.0001,
+      "loss": 2.9233,
+      "loss/crossentropy": 2.289682912826538,
+      "loss/hidden": 2.646875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2165387198328972,
+      "loss/reg": 0.0,
+      "step": 36090
+    },
+    {
+      "epoch": 0.2375,
+      "grad_norm": 1.9921875,
+      "grad_norm_var": 0.09746068318684896,
+      "learning_rate": 0.0001,
+      "loss": 3.0195,
+      "loss/crossentropy": 2.332055389881134,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21832392662763594,
+      "loss/reg": 0.0,
+      "step": 36100
+    },
+    {
+      "epoch": 0.2375657894736842,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.07844619750976563,
+      "learning_rate": 0.0001,
+      "loss": 3.0759,
+      "loss/crossentropy": 2.5219852209091185,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27026659697294236,
+      "loss/reg": 0.0,
+      "step": 36110
+    },
+    {
+      "epoch": 0.23763157894736842,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.07219950358072917,
+      "learning_rate": 0.0001,
+      "loss": 2.9762,
+      "loss/crossentropy": 2.2172631919384003,
+      "loss/hidden": 2.6953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22206666469573974,
+      "loss/reg": 0.0,
+      "step": 36120
+    },
+    {
+      "epoch": 0.23769736842105263,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.11806640625,
+      "learning_rate": 0.0001,
+      "loss": 2.9813,
+      "loss/crossentropy": 2.235606110095978,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23971866220235824,
+      "loss/reg": 0.0,
+      "step": 36130
+    },
+    {
+      "epoch": 0.23776315789473684,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 1.190087890625,
+      "learning_rate": 0.0001,
+      "loss": 3.0329,
+      "loss/crossentropy": 2.2865034997463227,
+      "loss/hidden": 2.821875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.32822101563215256,
+      "loss/reg": 0.0,
+      "step": 36140
+    },
+    {
+      "epoch": 0.23782894736842106,
+      "grad_norm": 2.625,
+      "grad_norm_var": 1.3279296875,
+      "learning_rate": 0.0001,
+      "loss": 3.0021,
+      "loss/crossentropy": 2.2038097441196443,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21367209404706955,
+      "loss/reg": 0.0,
+      "step": 36150
+    },
+    {
+      "epoch": 0.23789473684210527,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.05592041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0254,
+      "loss/crossentropy": 2.2345643639564514,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27571815252304077,
+      "loss/reg": 0.0,
+      "step": 36160
+    },
+    {
+      "epoch": 0.23796052631578948,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.11005757649739584,
+      "learning_rate": 0.0001,
+      "loss": 3.039,
+      "loss/crossentropy": 2.1099373579025267,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21341636329889296,
+      "loss/reg": 0.0,
+      "step": 36170
+    },
+    {
+      "epoch": 0.2380263157894737,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.114306640625,
+      "learning_rate": 0.0001,
+      "loss": 2.9229,
+      "loss/crossentropy": 2.1778077363967894,
+      "loss/hidden": 2.5546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18448112457990645,
+      "loss/reg": 0.0,
+      "step": 36180
+    },
+    {
+      "epoch": 0.23809210526315788,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.042867024739583336,
+      "learning_rate": 0.0001,
+      "loss": 2.9876,
+      "loss/crossentropy": 2.5516823649406435,
+      "loss/hidden": 2.7046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25054328292608263,
+      "loss/reg": 0.0,
+      "step": 36190
+    },
+    {
+      "epoch": 0.2381578947368421,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.07666727701822916,
+      "learning_rate": 0.0001,
+      "loss": 2.9734,
+      "loss/crossentropy": 2.257242572307587,
+      "loss/hidden": 2.6921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2207578793168068,
+      "loss/reg": 0.0,
+      "step": 36200
+    },
+    {
+      "epoch": 0.2382236842105263,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.07727864583333334,
+      "learning_rate": 0.0001,
+      "loss": 2.999,
+      "loss/crossentropy": 2.093051493167877,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2044597864151001,
+      "loss/reg": 0.0,
+      "step": 36210
+    },
+    {
+      "epoch": 0.23828947368421052,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.11516825358072917,
+      "learning_rate": 0.0001,
+      "loss": 2.9287,
+      "loss/crossentropy": 2.1220389723777773,
+      "loss/hidden": 2.6828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19910518899559976,
+      "loss/reg": 0.0,
+      "step": 36220
+    },
+    {
+      "epoch": 0.23835526315789474,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.2674763997395833,
+      "learning_rate": 0.0001,
+      "loss": 3.0577,
+      "loss/crossentropy": 2.1114853382110597,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2416158512234688,
+      "loss/reg": 0.0,
+      "step": 36230
+    },
+    {
+      "epoch": 0.23842105263157895,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.05162353515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9816,
+      "loss/crossentropy": 2.332682567834854,
+      "loss/hidden": 2.640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2057093556970358,
+      "loss/reg": 0.0,
+      "step": 36240
+    },
+    {
+      "epoch": 0.23848684210526316,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.04888916015625,
+      "learning_rate": 0.0001,
+      "loss": 2.9878,
+      "loss/crossentropy": 2.26646374464035,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22881892770528794,
+      "loss/reg": 0.0,
+      "step": 36250
+    },
+    {
+      "epoch": 0.23855263157894738,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.0685455322265625,
+      "learning_rate": 0.0001,
+      "loss": 3.041,
+      "loss/crossentropy": 2.0819905757904054,
+      "loss/hidden": 2.7578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2033102184534073,
+      "loss/reg": 0.0,
+      "step": 36260
+    },
+    {
+      "epoch": 0.2386184210526316,
+      "grad_norm": 3.453125,
+      "grad_norm_var": 2.275194295247396,
+      "learning_rate": 0.0001,
+      "loss": 3.0941,
+      "loss/crossentropy": 2.1767791748046874,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24875595569610595,
+      "loss/reg": 0.0,
+      "step": 36270
+    },
+    {
+      "epoch": 0.23868421052631578,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 3.0041412353515624,
+      "learning_rate": 0.0001,
+      "loss": 3.0775,
+      "loss/crossentropy": 2.389254295825958,
+      "loss/hidden": 2.659375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20708534121513367,
+      "loss/reg": 0.0,
+      "step": 36280
+    },
+    {
+      "epoch": 0.23875,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 1.19429931640625,
+      "learning_rate": 0.0001,
+      "loss": 2.9765,
+      "loss/crossentropy": 2.2551061868667603,
+      "loss/hidden": 2.8640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23167352080345155,
+      "loss/reg": 0.0,
+      "step": 36290
+    },
+    {
+      "epoch": 0.2388157894736842,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.4574127197265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0732,
+      "loss/crossentropy": 2.3170044660568236,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22510029524564742,
+      "loss/reg": 0.0,
+      "step": 36300
+    },
+    {
+      "epoch": 0.23888157894736842,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.6108904520670573,
+      "learning_rate": 0.0001,
+      "loss": 3.0454,
+      "loss/crossentropy": 1.9875805854797364,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21809106022119523,
+      "loss/reg": 0.0,
+      "step": 36310
+    },
+    {
+      "epoch": 0.23894736842105263,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.7914265950520833,
+      "learning_rate": 0.0001,
+      "loss": 2.9925,
+      "loss/crossentropy": 2.3829528450965882,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2381367027759552,
+      "loss/reg": 0.0,
+      "step": 36320
+    },
+    {
+      "epoch": 0.23901315789473684,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.3254778544108073,
+      "learning_rate": 0.0001,
+      "loss": 3.0144,
+      "loss/crossentropy": 2.196928286552429,
+      "loss/hidden": 2.8453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24151963144540786,
+      "loss/reg": 0.0,
+      "step": 36330
+    },
+    {
+      "epoch": 0.23907894736842106,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.0215240478515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9815,
+      "loss/crossentropy": 2.2826131939888,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23180599957704545,
+      "loss/reg": 0.0,
+      "step": 36340
+    },
+    {
+      "epoch": 0.23914473684210527,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.7601847330729167,
+      "learning_rate": 0.0001,
+      "loss": 3.025,
+      "loss/crossentropy": 2.4469813466072083,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21003625690937042,
+      "loss/reg": 0.0,
+      "step": 36350
+    },
+    {
+      "epoch": 0.23921052631578948,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.04129231770833333,
+      "learning_rate": 0.0001,
+      "loss": 2.9786,
+      "loss/crossentropy": 2.5087629079818727,
+      "loss/hidden": 2.7390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22830777615308762,
+      "loss/reg": 0.0,
+      "step": 36360
+    },
+    {
+      "epoch": 0.2392763157894737,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.05718994140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0514,
+      "loss/crossentropy": 2.0158798813819887,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22552752420306205,
+      "loss/reg": 0.0,
+      "step": 36370
+    },
+    {
+      "epoch": 0.23934210526315788,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.056298828125,
+      "learning_rate": 0.0001,
+      "loss": 3.0251,
+      "loss/crossentropy": 1.9761714577674865,
+      "loss/hidden": 2.859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23326522260904312,
+      "loss/reg": 0.0,
+      "step": 36380
+    },
+    {
+      "epoch": 0.2394078947368421,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.19085286458333334,
+      "learning_rate": 0.0001,
+      "loss": 3.0543,
+      "loss/crossentropy": 2.1724894046783447,
+      "loss/hidden": 2.8859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2123780742287636,
+      "loss/reg": 0.0,
+      "step": 36390
+    },
+    {
+      "epoch": 0.2394736842105263,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.1413726806640625,
+      "learning_rate": 0.0001,
+      "loss": 2.9757,
+      "loss/crossentropy": 2.1725098967552183,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2148168534040451,
+      "loss/reg": 0.0,
+      "step": 36400
+    },
+    {
+      "epoch": 0.23953947368421052,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 4.847835032145182,
+      "learning_rate": 0.0001,
+      "loss": 3.046,
+      "loss/crossentropy": 2.4329644799232484,
+      "loss/hidden": 2.8671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26010534912347794,
+      "loss/reg": 0.0,
+      "step": 36410
+    },
+    {
+      "epoch": 0.23960526315789474,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 4.855410766601563,
+      "learning_rate": 0.0001,
+      "loss": 2.9563,
+      "loss/crossentropy": 2.215455192327499,
+      "loss/hidden": 2.909375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23449745997786522,
+      "loss/reg": 0.0,
+      "step": 36420
+    },
+    {
+      "epoch": 0.23967105263157895,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.0796539306640625,
+      "learning_rate": 0.0001,
+      "loss": 3.015,
+      "loss/crossentropy": 2.411003601551056,
+      "loss/hidden": 2.8921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27160040885210035,
+      "loss/reg": 0.0,
+      "step": 36430
+    },
+    {
+      "epoch": 0.23973684210526316,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.0956207275390625,
+      "learning_rate": 0.0001,
+      "loss": 2.9926,
+      "loss/crossentropy": 2.542527449131012,
+      "loss/hidden": 2.8,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23901409804821014,
+      "loss/reg": 0.0,
+      "step": 36440
+    },
+    {
+      "epoch": 0.23980263157894738,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.1344146728515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0325,
+      "loss/crossentropy": 2.28466220498085,
+      "loss/hidden": 2.665625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21214617267251015,
+      "loss/reg": 0.0,
+      "step": 36450
+    },
+    {
+      "epoch": 0.2398684210526316,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.07780659993489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0155,
+      "loss/crossentropy": 2.521458077430725,
+      "loss/hidden": 2.7625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24949344843626023,
+      "loss/reg": 0.0,
+      "step": 36460
+    },
+    {
+      "epoch": 0.23993421052631578,
+      "grad_norm": 3.171875,
+      "grad_norm_var": 0.10063374837239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0598,
+      "loss/crossentropy": 2.11941602230072,
+      "loss/hidden": 2.8125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.265853063762188,
+      "loss/reg": 0.0,
+      "step": 36470
+    },
+    {
+      "epoch": 0.24,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.15134175618489584,
+      "learning_rate": 0.0001,
+      "loss": 3.1036,
+      "loss/crossentropy": 2.159947466850281,
+      "loss/hidden": 2.7875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23642594143748283,
+      "loss/reg": 0.0,
+      "step": 36480
+    },
+    {
+      "epoch": 0.2400657894736842,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.1106109619140625,
+      "learning_rate": 0.0001,
+      "loss": 2.9888,
+      "loss/crossentropy": 2.2564534187316894,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24869659841060637,
+      "loss/reg": 0.0,
+      "step": 36490
+    },
+    {
+      "epoch": 0.24013157894736842,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.04983723958333333,
+      "learning_rate": 0.0001,
+      "loss": 3.107,
+      "loss/crossentropy": 2.3479647517204283,
+      "loss/hidden": 2.715625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2286713719367981,
+      "loss/reg": 0.0,
+      "step": 36500
+    },
+    {
+      "epoch": 0.24019736842105263,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.07700093587239583,
+      "learning_rate": 0.0001,
+      "loss": 3.0664,
+      "loss/crossentropy": 2.4374531388282774,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23715483397245407,
+      "loss/reg": 0.0,
+      "step": 36510
+    },
+    {
+      "epoch": 0.24026315789473685,
+      "grad_norm": 3.171875,
+      "grad_norm_var": 0.28349202473958335,
+      "learning_rate": 0.0001,
+      "loss": 3.0541,
+      "loss/crossentropy": 1.9271621584892273,
+      "loss/hidden": 2.91875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2382538564503193,
+      "loss/reg": 0.0,
+      "step": 36520
+    },
+    {
+      "epoch": 0.24032894736842106,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.26936747233072916,
+      "learning_rate": 0.0001,
+      "loss": 2.9974,
+      "loss/crossentropy": 2.3039157390594482,
+      "loss/hidden": 2.5375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18885463178157808,
+      "loss/reg": 0.0,
+      "step": 36530
+    },
+    {
+      "epoch": 0.24039473684210527,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.04563700358072917,
+      "learning_rate": 0.0001,
+      "loss": 2.9944,
+      "loss/crossentropy": 2.1725877285003663,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22165239751338958,
+      "loss/reg": 0.0,
+      "step": 36540
+    },
+    {
+      "epoch": 0.24046052631578949,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.026154581705729166,
+      "learning_rate": 0.0001,
+      "loss": 3.011,
+      "loss/crossentropy": 2.393988037109375,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25498510152101517,
+      "loss/reg": 0.0,
+      "step": 36550
+    },
+    {
+      "epoch": 0.24052631578947367,
+      "grad_norm": 5.125,
+      "grad_norm_var": 0.5747548421223958,
+      "learning_rate": 0.0001,
+      "loss": 3.0477,
+      "loss/crossentropy": 2.3300102829933165,
+      "loss/hidden": 2.921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2910366252064705,
+      "loss/reg": 0.0,
+      "step": 36560
+    },
+    {
+      "epoch": 0.24059210526315788,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.541943359375,
+      "learning_rate": 0.0001,
+      "loss": 3.0418,
+      "loss/crossentropy": 2.2268447399139406,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.236025869846344,
+      "loss/reg": 0.0,
+      "step": 36570
+    },
+    {
+      "epoch": 0.2406578947368421,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.11101786295572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0594,
+      "loss/crossentropy": 2.4716415405273438,
+      "loss/hidden": 2.8453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2721579551696777,
+      "loss/reg": 0.0,
+      "step": 36580
+    },
+    {
+      "epoch": 0.2407236842105263,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.09099934895833334,
+      "learning_rate": 0.0001,
+      "loss": 2.9772,
+      "loss/crossentropy": 2.2797286033630373,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24081944078207015,
+      "loss/reg": 0.0,
+      "step": 36590
+    },
+    {
+      "epoch": 0.24078947368421053,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.04234110514322917,
+      "learning_rate": 0.0001,
+      "loss": 2.9358,
+      "loss/crossentropy": 2.2048792719841,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2422194480895996,
+      "loss/reg": 0.0,
+      "step": 36600
+    },
+    {
+      "epoch": 0.24085526315789474,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.061766560872395834,
+      "learning_rate": 0.0001,
+      "loss": 3.0394,
+      "loss/crossentropy": 2.4691983580589296,
+      "loss/hidden": 2.903125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2871833577752113,
+      "loss/reg": 0.0,
+      "step": 36610
+    },
+    {
+      "epoch": 0.24092105263157895,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.07330322265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0204,
+      "loss/crossentropy": 2.3721364736557007,
+      "loss/hidden": 2.7421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24769402742385865,
+      "loss/reg": 0.0,
+      "step": 36620
+    },
+    {
+      "epoch": 0.24098684210526317,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.08896077473958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0345,
+      "loss/crossentropy": 2.6213492870330812,
+      "loss/hidden": 2.6671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22394069284200668,
+      "loss/reg": 0.0,
+      "step": 36630
+    },
+    {
+      "epoch": 0.24105263157894738,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.0711822509765625,
+      "learning_rate": 0.0001,
+      "loss": 2.9613,
+      "loss/crossentropy": 2.4529595017433166,
+      "loss/hidden": 2.696875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2483013778924942,
+      "loss/reg": 0.0,
+      "step": 36640
+    },
+    {
+      "epoch": 0.24111842105263157,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.024247233072916666,
+      "learning_rate": 0.0001,
+      "loss": 3.0826,
+      "loss/crossentropy": 2.5188237547874452,
+      "loss/hidden": 2.809375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25313877910375593,
+      "loss/reg": 0.0,
+      "step": 36650
+    },
+    {
+      "epoch": 0.24118421052631578,
+      "grad_norm": 1.96875,
+      "grad_norm_var": 0.41513264973958336,
+      "learning_rate": 0.0001,
+      "loss": 3.0485,
+      "loss/crossentropy": 2.358896279335022,
+      "loss/hidden": 2.8015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2380443513393402,
+      "loss/reg": 0.0,
+      "step": 36660
+    },
+    {
+      "epoch": 0.24125,
+      "grad_norm": 2088763392.0,
+      "grad_norm_var": 4.570212428561056e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.3122,
+      "loss/crossentropy": 2.4535842657089235,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23562741428613662,
+      "loss/reg": 0.0,
+      "step": 36670
+    },
+    {
+      "epoch": 0.2413157894736842,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 4.5702124295693926e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0286,
+      "loss/crossentropy": 2.1914322853088377,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25221034586429597,
+      "loss/reg": 0.0,
+      "step": 36680
+    },
+    {
+      "epoch": 0.24138157894736842,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.031927235921223956,
+      "learning_rate": 0.0001,
+      "loss": 3.0104,
+      "loss/crossentropy": 2.4763585209846495,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2514804720878601,
+      "loss/reg": 0.0,
+      "step": 36690
+    },
+    {
+      "epoch": 0.24144736842105263,
+      "grad_norm": 3.1875,
+      "grad_norm_var": 0.08718236287434895,
+      "learning_rate": 0.0001,
+      "loss": 3.159,
+      "loss/crossentropy": 2.0223356008529665,
+      "loss/hidden": 3.146875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26726412028074265,
+      "loss/reg": 0.0,
+      "step": 36700
+    },
+    {
+      "epoch": 0.24151315789473685,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.056004842122395836,
+      "learning_rate": 0.0001,
+      "loss": 3.0129,
+      "loss/crossentropy": 2.2822505116462706,
+      "loss/hidden": 2.696875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21236415654420854,
+      "loss/reg": 0.0,
+      "step": 36710
+    },
+    {
+      "epoch": 0.24157894736842106,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.049605305989583334,
+      "learning_rate": 0.0001,
+      "loss": 2.9728,
+      "loss/crossentropy": 2.152122360467911,
+      "loss/hidden": 2.765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22159827202558519,
+      "loss/reg": 0.0,
+      "step": 36720
+    },
+    {
+      "epoch": 0.24164473684210527,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.2835845947265625,
+      "learning_rate": 0.0001,
+      "loss": 3.1062,
+      "loss/crossentropy": 2.415015733242035,
+      "loss/hidden": 2.765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2637294501066208,
+      "loss/reg": 0.0,
+      "step": 36730
+    },
+    {
+      "epoch": 0.2417105263157895,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.24265950520833332,
+      "learning_rate": 0.0001,
+      "loss": 3.0508,
+      "loss/crossentropy": 2.143397808074951,
+      "loss/hidden": 2.81875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26159815937280656,
+      "loss/reg": 0.0,
+      "step": 36740
+    },
+    {
+      "epoch": 0.24177631578947367,
+      "grad_norm": 2.875,
+      "grad_norm_var": 0.0584381103515625,
+      "learning_rate": 0.0001,
+      "loss": 3.1012,
+      "loss/crossentropy": 2.3698354959487915,
+      "loss/hidden": 2.8609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26768290251493454,
+      "loss/reg": 0.0,
+      "step": 36750
+    },
+    {
+      "epoch": 0.24184210526315789,
+      "grad_norm": 1.9375,
+      "grad_norm_var": 0.9022532145182292,
+      "learning_rate": 0.0001,
+      "loss": 2.9442,
+      "loss/crossentropy": 2.1970738768577576,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2042816124856472,
+      "loss/reg": 0.0,
+      "step": 36760
+    },
+    {
+      "epoch": 0.2419078947368421,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.9153279622395833,
+      "learning_rate": 0.0001,
+      "loss": 3.0157,
+      "loss/crossentropy": 2.345754420757294,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23231206387281417,
+      "loss/reg": 0.0,
+      "step": 36770
+    },
+    {
+      "epoch": 0.2419736842105263,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.08898111979166666,
+      "learning_rate": 0.0001,
+      "loss": 3.0409,
+      "loss/crossentropy": 2.3259128451347353,
+      "loss/hidden": 2.9046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23587240129709244,
+      "loss/reg": 0.0,
+      "step": 36780
+    },
+    {
+      "epoch": 0.24203947368421053,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.07978108723958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0165,
+      "loss/crossentropy": 2.5162180185317995,
+      "loss/hidden": 2.7515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2471260607242584,
+      "loss/reg": 0.0,
+      "step": 36790
+    },
+    {
+      "epoch": 0.24210526315789474,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.03634440104166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0087,
+      "loss/crossentropy": 2.3737236499786376,
+      "loss/hidden": 2.6671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2186498686671257,
+      "loss/reg": 0.0,
+      "step": 36800
+    },
+    {
+      "epoch": 0.24217105263157895,
+      "grad_norm": 2.09375,
+      "grad_norm_var": 0.049605305989583334,
+      "learning_rate": 0.0001,
+      "loss": 3.0755,
+      "loss/crossentropy": 2.353866970539093,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22997004687786102,
+      "loss/reg": 0.0,
+      "step": 36810
+    },
+    {
+      "epoch": 0.24223684210526317,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.11220296223958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0278,
+      "loss/crossentropy": 2.421842908859253,
+      "loss/hidden": 2.790625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21242005676031112,
+      "loss/reg": 0.0,
+      "step": 36820
+    },
+    {
+      "epoch": 0.24230263157894738,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.16858622233072917,
+      "learning_rate": 0.0001,
+      "loss": 3.1032,
+      "loss/crossentropy": 2.3163668155670165,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24295762777328492,
+      "loss/reg": 0.0,
+      "step": 36830
+    },
+    {
+      "epoch": 0.24236842105263157,
+      "grad_norm": 2.984375,
+      "grad_norm_var": 0.0773150126139323,
+      "learning_rate": 0.0001,
+      "loss": 2.9972,
+      "loss/crossentropy": 2.162323606014252,
+      "loss/hidden": 2.9359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2663910940289497,
+      "loss/reg": 0.0,
+      "step": 36840
+    },
+    {
+      "epoch": 0.24243421052631578,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.21162923177083334,
+      "learning_rate": 0.0001,
+      "loss": 2.962,
+      "loss/crossentropy": 2.159565594792366,
+      "loss/hidden": 2.6578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19483270347118378,
+      "loss/reg": 0.0,
+      "step": 36850
+    },
+    {
+      "epoch": 0.2425,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.3277259826660156,
+      "learning_rate": 0.0001,
+      "loss": 3.0165,
+      "loss/crossentropy": 2.5494640946388243,
+      "loss/hidden": 2.709375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21899326890707016,
+      "loss/reg": 0.0,
+      "step": 36860
+    },
+    {
+      "epoch": 0.2425657894736842,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.02713190714518229,
+      "learning_rate": 0.0001,
+      "loss": 2.9519,
+      "loss/crossentropy": 2.3883594393730165,
+      "loss/hidden": 2.7453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21375252604484557,
+      "loss/reg": 0.0,
+      "step": 36870
+    },
+    {
+      "epoch": 0.24263157894736842,
+      "grad_norm": 1.9296875,
+      "grad_norm_var": 0.03628107706705729,
+      "learning_rate": 0.0001,
+      "loss": 2.9473,
+      "loss/crossentropy": 2.1211004137992857,
+      "loss/hidden": 2.8984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2303142726421356,
+      "loss/reg": 0.0,
+      "step": 36880
+    },
+    {
+      "epoch": 0.24269736842105263,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.053929646809895836,
+      "learning_rate": 0.0001,
+      "loss": 2.9946,
+      "loss/crossentropy": 2.2483023762702943,
+      "loss/hidden": 2.965625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2498287908732891,
+      "loss/reg": 0.0,
+      "step": 36890
+    },
+    {
+      "epoch": 0.24276315789473685,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.026569620768229166,
+      "learning_rate": 0.0001,
+      "loss": 3.0207,
+      "loss/crossentropy": 2.4569414138793944,
+      "loss/hidden": 2.703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22985844165086747,
+      "loss/reg": 0.0,
+      "step": 36900
+    },
+    {
+      "epoch": 0.24282894736842106,
+      "grad_norm": 2.953125,
+      "grad_norm_var": 0.06067708333333333,
+      "learning_rate": 0.0001,
+      "loss": 3.0104,
+      "loss/crossentropy": 2.438031816482544,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.207527032494545,
+      "loss/reg": 0.0,
+      "step": 36910
+    },
+    {
+      "epoch": 0.24289473684210527,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.08677469889322917,
+      "learning_rate": 0.0001,
+      "loss": 2.9992,
+      "loss/crossentropy": 2.4970327615737915,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24948519468307495,
+      "loss/reg": 0.0,
+      "step": 36920
+    },
+    {
+      "epoch": 0.24296052631578946,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.07526041666666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0188,
+      "loss/crossentropy": 2.4824806571006777,
+      "loss/hidden": 2.9140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26679628640413283,
+      "loss/reg": 0.0,
+      "step": 36930
+    },
+    {
+      "epoch": 0.24302631578947367,
+      "grad_norm": 3.125,
+      "grad_norm_var": 0.15319595336914063,
+      "learning_rate": 0.0001,
+      "loss": 3.0338,
+      "loss/crossentropy": 2.588907778263092,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2500623852014542,
+      "loss/reg": 0.0,
+      "step": 36940
+    },
+    {
+      "epoch": 0.2430921052631579,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.105810546875,
+      "learning_rate": 0.0001,
+      "loss": 3.0026,
+      "loss/crossentropy": 2.233619010448456,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23702021837234497,
+      "loss/reg": 0.0,
+      "step": 36950
+    },
+    {
+      "epoch": 0.2431578947368421,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.032210286458333334,
+      "learning_rate": 0.0001,
+      "loss": 2.9011,
+      "loss/crossentropy": 2.494797945022583,
+      "loss/hidden": 2.65,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2184271454811096,
+      "loss/reg": 0.0,
+      "step": 36960
+    },
+    {
+      "epoch": 0.24322368421052631,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.026590983072916668,
+      "learning_rate": 0.0001,
+      "loss": 3.048,
+      "loss/crossentropy": 2.323407733440399,
+      "loss/hidden": 2.8640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2522672712802887,
+      "loss/reg": 0.0,
+      "step": 36970
+    },
+    {
+      "epoch": 0.24328947368421053,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.04624735514322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0306,
+      "loss/crossentropy": 2.248076152801514,
+      "loss/hidden": 2.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22199858874082565,
+      "loss/reg": 0.0,
+      "step": 36980
+    },
+    {
+      "epoch": 0.24335526315789474,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.38259989420572915,
+      "learning_rate": 0.0001,
+      "loss": 2.9916,
+      "loss/crossentropy": 2.251713329553604,
+      "loss/hidden": 2.6640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21513912975788116,
+      "loss/reg": 0.0,
+      "step": 36990
+    },
+    {
+      "epoch": 0.24342105263157895,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.384716796875,
+      "learning_rate": 0.0001,
+      "loss": 3.0287,
+      "loss/crossentropy": 2.0817331850528715,
+      "loss/hidden": 2.8484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21513576656579972,
+      "loss/reg": 0.0,
+      "step": 37000
+    },
+    {
+      "epoch": 0.24348684210526317,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.07632420857747396,
+      "learning_rate": 0.0001,
+      "loss": 3.0023,
+      "loss/crossentropy": 2.351958858966827,
+      "loss/hidden": 3.0078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27439084053039553,
+      "loss/reg": 0.0,
+      "step": 37010
+    },
+    {
+      "epoch": 0.24355263157894738,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.12300186157226563,
+      "learning_rate": 0.0001,
+      "loss": 3.0848,
+      "loss/crossentropy": 2.2964416265487673,
+      "loss/hidden": 2.76875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28409299105405805,
+      "loss/reg": 0.0,
+      "step": 37020
+    },
+    {
+      "epoch": 0.24361842105263157,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.18337173461914064,
+      "learning_rate": 0.0001,
+      "loss": 2.9231,
+      "loss/crossentropy": 2.2850769579410555,
+      "loss/hidden": 2.6234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20450889468193054,
+      "loss/reg": 0.0,
+      "step": 37030
+    },
+    {
+      "epoch": 0.24368421052631578,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.18819071451822916,
+      "learning_rate": 0.0001,
+      "loss": 3.0757,
+      "loss/crossentropy": 2.157426190376282,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23339182287454605,
+      "loss/reg": 0.0,
+      "step": 37040
+    },
+    {
+      "epoch": 0.24375,
+      "grad_norm": 2.796875,
+      "grad_norm_var": 0.0508209228515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0667,
+      "loss/crossentropy": 2.3142905950546266,
+      "loss/hidden": 2.91875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27201273292303085,
+      "loss/reg": 0.0,
+      "step": 37050
+    },
+    {
+      "epoch": 0.2438157894736842,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.054230753580729166,
+      "learning_rate": 0.0001,
+      "loss": 2.9379,
+      "loss/crossentropy": 2.182029736042023,
+      "loss/hidden": 2.765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2124377816915512,
+      "loss/reg": 0.0,
+      "step": 37060
+    },
+    {
+      "epoch": 0.24388157894736842,
+      "grad_norm": 1.96875,
+      "grad_norm_var": 0.05328776041666667,
+      "learning_rate": 0.0001,
+      "loss": 2.9728,
+      "loss/crossentropy": 2.47544287443161,
+      "loss/hidden": 2.734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20328531116247178,
+      "loss/reg": 0.0,
+      "step": 37070
+    },
+    {
+      "epoch": 0.24394736842105263,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.11983617146809895,
+      "learning_rate": 0.0001,
+      "loss": 3.038,
+      "loss/crossentropy": 2.315932643413544,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2507309168577194,
+      "loss/reg": 0.0,
+      "step": 37080
+    },
+    {
+      "epoch": 0.24401315789473685,
+      "grad_norm": 3.125,
+      "grad_norm_var": 0.5823707580566406,
+      "learning_rate": 0.0001,
+      "loss": 3.0428,
+      "loss/crossentropy": 2.0966883838176726,
+      "loss/hidden": 2.9171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24450836032629014,
+      "loss/reg": 0.0,
+      "step": 37090
+    },
+    {
+      "epoch": 0.24407894736842106,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.562109375,
+      "learning_rate": 0.0001,
+      "loss": 3.0281,
+      "loss/crossentropy": 2.111643207073212,
+      "loss/hidden": 2.8578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22489524409174919,
+      "loss/reg": 0.0,
+      "step": 37100
+    },
+    {
+      "epoch": 0.24414473684210528,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.048238118489583336,
+      "learning_rate": 0.0001,
+      "loss": 2.9667,
+      "loss/crossentropy": 2.2703915894031526,
+      "loss/hidden": 2.7421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20935535281896592,
+      "loss/reg": 0.0,
+      "step": 37110
+    },
+    {
+      "epoch": 0.24421052631578946,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.12493082682291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0306,
+      "loss/crossentropy": 2.4712467312812807,
+      "loss/hidden": 2.653125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21050616949796677,
+      "loss/reg": 0.0,
+      "step": 37120
+    },
+    {
+      "epoch": 0.24427631578947367,
+      "grad_norm": 2.078125,
+      "grad_norm_var": 0.22398173014322917,
+      "learning_rate": 0.0001,
+      "loss": 2.9499,
+      "loss/crossentropy": 2.1810320615768433,
+      "loss/hidden": 2.6078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1909588247537613,
+      "loss/reg": 0.0,
+      "step": 37130
+    },
+    {
+      "epoch": 0.2443421052631579,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.18337300618489583,
+      "learning_rate": 0.0001,
+      "loss": 3.0502,
+      "loss/crossentropy": 2.259117543697357,
+      "loss/hidden": 2.8859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2917395427823067,
+      "loss/reg": 0.0,
+      "step": 37140
+    },
+    {
+      "epoch": 0.2444078947368421,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.04478251139322917,
+      "learning_rate": 0.0001,
+      "loss": 3.003,
+      "loss/crossentropy": 2.3330495953559875,
+      "loss/hidden": 2.825,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25376534312963483,
+      "loss/reg": 0.0,
+      "step": 37150
+    },
+    {
+      "epoch": 0.24447368421052632,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.025809733072916667,
+      "learning_rate": 0.0001,
+      "loss": 2.9673,
+      "loss/crossentropy": 2.2595450043678285,
+      "loss/hidden": 2.815625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24186031073331832,
+      "loss/reg": 0.0,
+      "step": 37160
+    },
+    {
+      "epoch": 0.24453947368421053,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.040257771809895836,
+      "learning_rate": 0.0001,
+      "loss": 3.0137,
+      "loss/crossentropy": 2.3244964241981507,
+      "loss/hidden": 2.6875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22835270911455155,
+      "loss/reg": 0.0,
+      "step": 37170
+    },
+    {
+      "epoch": 0.24460526315789474,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.18911844889322918,
+      "learning_rate": 0.0001,
+      "loss": 3.095,
+      "loss/crossentropy": 2.1679759979248048,
+      "loss/hidden": 2.7515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22593683898448944,
+      "loss/reg": 0.0,
+      "step": 37180
+    },
+    {
+      "epoch": 0.24467105263157896,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.0819488525390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0209,
+      "loss/crossentropy": 2.3522397756576536,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22514844089746475,
+      "loss/reg": 0.0,
+      "step": 37190
+    },
+    {
+      "epoch": 0.24473684210526317,
+      "grad_norm": 3.171875,
+      "grad_norm_var": 0.14850972493489584,
+      "learning_rate": 0.0001,
+      "loss": 3.0028,
+      "loss/crossentropy": 2.1798877120018005,
+      "loss/hidden": 2.84375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26098706275224687,
+      "loss/reg": 0.0,
+      "step": 37200
+    },
+    {
+      "epoch": 0.24480263157894736,
+      "grad_norm": 3.4375,
+      "grad_norm_var": 0.16476949055989584,
+      "learning_rate": 0.0001,
+      "loss": 3.0878,
+      "loss/crossentropy": 2.3389397978782656,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23207588642835617,
+      "loss/reg": 0.0,
+      "step": 37210
+    },
+    {
+      "epoch": 0.24486842105263157,
+      "grad_norm": 1.96875,
+      "grad_norm_var": 0.13612874348958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0629,
+      "loss/crossentropy": 2.29145188331604,
+      "loss/hidden": 2.9859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23295368999242783,
+      "loss/reg": 0.0,
+      "step": 37220
+    },
+    {
+      "epoch": 0.24493421052631578,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.0704742431640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0207,
+      "loss/crossentropy": 2.263294315338135,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21528264433145522,
+      "loss/reg": 0.0,
+      "step": 37230
+    },
+    {
+      "epoch": 0.245,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.04729410807291667,
+      "learning_rate": 0.0001,
+      "loss": 2.9616,
+      "loss/crossentropy": 2.4668697357177733,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2833425417542458,
+      "loss/reg": 0.0,
+      "step": 37240
+    },
+    {
+      "epoch": 0.2450657894736842,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.1173736572265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0141,
+      "loss/crossentropy": 2.310600745677948,
+      "loss/hidden": 2.6984375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2221137210726738,
+      "loss/reg": 0.0,
+      "step": 37250
+    },
+    {
+      "epoch": 0.24513157894736842,
+      "grad_norm": 3.359375,
+      "grad_norm_var": 0.15969136555989583,
+      "learning_rate": 0.0001,
+      "loss": 3.0175,
+      "loss/crossentropy": 2.113830578327179,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2413550451397896,
+      "loss/reg": 0.0,
+      "step": 37260
+    },
+    {
+      "epoch": 0.24519736842105264,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.12274983723958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0058,
+      "loss/crossentropy": 1.9976770758628846,
+      "loss/hidden": 2.7171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19502876847982406,
+      "loss/reg": 0.0,
+      "step": 37270
+    },
+    {
+      "epoch": 0.24526315789473685,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.11809488932291666,
+      "learning_rate": 0.0001,
+      "loss": 2.9752,
+      "loss/crossentropy": 2.1309264838695525,
+      "loss/hidden": 2.65,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20758389085531234,
+      "loss/reg": 0.0,
+      "step": 37280
+    },
+    {
+      "epoch": 0.24532894736842106,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.12764383951822916,
+      "learning_rate": 0.0001,
+      "loss": 3.044,
+      "loss/crossentropy": 2.142688637971878,
+      "loss/hidden": 2.75625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2955352425575256,
+      "loss/reg": 0.0,
+      "step": 37290
+    },
+    {
+      "epoch": 0.24539473684210528,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.0466949462890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9294,
+      "loss/crossentropy": 2.255223333835602,
+      "loss/hidden": 2.753125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22907140627503395,
+      "loss/reg": 0.0,
+      "step": 37300
+    },
+    {
+      "epoch": 0.24546052631578946,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.085693359375,
+      "learning_rate": 0.0001,
+      "loss": 2.9874,
+      "loss/crossentropy": 2.121692883968353,
+      "loss/hidden": 2.7140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.195634426176548,
+      "loss/reg": 0.0,
+      "step": 37310
+    },
+    {
+      "epoch": 0.24552631578947368,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.12754694620768228,
+      "learning_rate": 0.0001,
+      "loss": 3.0103,
+      "loss/crossentropy": 2.2369035363197325,
+      "loss/hidden": 2.8828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2212720662355423,
+      "loss/reg": 0.0,
+      "step": 37320
+    },
+    {
+      "epoch": 0.2455921052631579,
+      "grad_norm": 2.953125,
+      "grad_norm_var": 0.14957046508789062,
+      "learning_rate": 0.0001,
+      "loss": 2.9211,
+      "loss/crossentropy": 1.6413456916809082,
+      "loss/hidden": 2.603125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18164349719882011,
+      "loss/reg": 0.0,
+      "step": 37330
+    },
+    {
+      "epoch": 0.2456578947368421,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.17405497233072917,
+      "learning_rate": 0.0001,
+      "loss": 3.0623,
+      "loss/crossentropy": 2.128761112689972,
+      "loss/hidden": 2.8359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20798836946487426,
+      "loss/reg": 0.0,
+      "step": 37340
+    },
+    {
+      "epoch": 0.24572368421052632,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.05827534993489583,
+      "learning_rate": 0.0001,
+      "loss": 2.9564,
+      "loss/crossentropy": 2.271031451225281,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21863894239068032,
+      "loss/reg": 0.0,
+      "step": 37350
+    },
+    {
+      "epoch": 0.24578947368421053,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.10803629557291666,
+      "learning_rate": 0.0001,
+      "loss": 3.0487,
+      "loss/crossentropy": 2.5324214935302733,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2654300585389137,
+      "loss/reg": 0.0,
+      "step": 37360
+    },
+    {
+      "epoch": 0.24585526315789474,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.0805816650390625,
+      "learning_rate": 0.0001,
+      "loss": 2.9872,
+      "loss/crossentropy": 2.476090502738953,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22690275460481643,
+      "loss/reg": 0.0,
+      "step": 37370
+    },
+    {
+      "epoch": 0.24592105263157896,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.0459136962890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9355,
+      "loss/crossentropy": 2.158538544178009,
+      "loss/hidden": 3.05625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2366640105843544,
+      "loss/reg": 0.0,
+      "step": 37380
+    },
+    {
+      "epoch": 0.24598684210526317,
+      "grad_norm": 1.8828125,
+      "grad_norm_var": 0.048860422770182294,
+      "learning_rate": 0.0001,
+      "loss": 3.0021,
+      "loss/crossentropy": 2.2990810751914976,
+      "loss/hidden": 2.690625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20744879692792892,
+      "loss/reg": 0.0,
+      "step": 37390
+    },
+    {
+      "epoch": 0.24605263157894736,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.052247873942057294,
+      "learning_rate": 0.0001,
+      "loss": 2.9696,
+      "loss/crossentropy": 2.2528524160385133,
+      "loss/hidden": 2.7546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22408121079206467,
+      "loss/reg": 0.0,
+      "step": 37400
+    },
+    {
+      "epoch": 0.24611842105263157,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.046122233072916664,
+      "learning_rate": 0.0001,
+      "loss": 2.991,
+      "loss/crossentropy": 2.4535847425460817,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27996278256177903,
+      "loss/reg": 0.0,
+      "step": 37410
+    },
+    {
+      "epoch": 0.24618421052631578,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.50777587890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9405,
+      "loss/crossentropy": 2.2263678312301636,
+      "loss/hidden": 2.6703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21364209055900574,
+      "loss/reg": 0.0,
+      "step": 37420
+    },
+    {
+      "epoch": 0.24625,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.0452301025390625,
+      "learning_rate": 0.0001,
+      "loss": 2.966,
+      "loss/crossentropy": 2.3186843156814576,
+      "loss/hidden": 2.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.251382140815258,
+      "loss/reg": 0.0,
+      "step": 37430
+    },
+    {
+      "epoch": 0.2463157894736842,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.035674794514973955,
+      "learning_rate": 0.0001,
+      "loss": 2.9334,
+      "loss/crossentropy": 2.3233260989189146,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26353937536478045,
+      "loss/reg": 0.0,
+      "step": 37440
+    },
+    {
+      "epoch": 0.24638157894736842,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.09699071248372396,
+      "learning_rate": 0.0001,
+      "loss": 3.0099,
+      "loss/crossentropy": 2.1750458002090456,
+      "loss/hidden": 2.725,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22462447360157967,
+      "loss/reg": 0.0,
+      "step": 37450
+    },
+    {
+      "epoch": 0.24644736842105264,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.09807942708333334,
+      "learning_rate": 0.0001,
+      "loss": 2.9668,
+      "loss/crossentropy": 2.59984233379364,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25129624158144,
+      "loss/reg": 0.0,
+      "step": 37460
+    },
+    {
+      "epoch": 0.24651315789473685,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.04807840983072917,
+      "learning_rate": 0.0001,
+      "loss": 2.9464,
+      "loss/crossentropy": 2.4232494115829466,
+      "loss/hidden": 2.6578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22763155549764633,
+      "loss/reg": 0.0,
+      "step": 37470
+    },
+    {
+      "epoch": 0.24657894736842106,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.2599283854166667,
+      "learning_rate": 0.0001,
+      "loss": 3.0313,
+      "loss/crossentropy": 2.314221677184105,
+      "loss/hidden": 2.775,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19144997373223305,
+      "loss/reg": 0.0,
+      "step": 37480
+    },
+    {
+      "epoch": 0.24664473684210525,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.3412068684895833,
+      "learning_rate": 0.0001,
+      "loss": 2.98,
+      "loss/crossentropy": 2.4459351778030394,
+      "loss/hidden": 2.8359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24818085730075837,
+      "loss/reg": 0.0,
+      "step": 37490
+    },
+    {
+      "epoch": 0.24671052631578946,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.31691080729166665,
+      "learning_rate": 0.0001,
+      "loss": 3.0414,
+      "loss/crossentropy": 2.1927252769470216,
+      "loss/hidden": 2.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2274562269449234,
+      "loss/reg": 0.0,
+      "step": 37500
+    },
+    {
+      "epoch": 0.24677631578947368,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.2977935791015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0205,
+      "loss/crossentropy": 2.0236656427383424,
+      "loss/hidden": 2.8734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23172616437077523,
+      "loss/reg": 0.0,
+      "step": 37510
+    },
+    {
+      "epoch": 0.2468421052631579,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.050593058268229164,
+      "learning_rate": 0.0001,
+      "loss": 2.9751,
+      "loss/crossentropy": 2.0957834839820864,
+      "loss/hidden": 2.8515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23045892268419266,
+      "loss/reg": 0.0,
+      "step": 37520
+    },
+    {
+      "epoch": 0.2469078947368421,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.09323628743489583,
+      "learning_rate": 0.0001,
+      "loss": 2.9756,
+      "loss/crossentropy": 2.189413595199585,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22707376182079314,
+      "loss/reg": 0.0,
+      "step": 37530
+    },
+    {
+      "epoch": 0.24697368421052632,
+      "grad_norm": 2.0,
+      "grad_norm_var": 0.056183878580729166,
+      "learning_rate": 0.0001,
+      "loss": 2.9589,
+      "loss/crossentropy": 2.733688974380493,
+      "loss/hidden": 2.709375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2146020546555519,
+      "loss/reg": 0.0,
+      "step": 37540
+    },
+    {
+      "epoch": 0.24703947368421053,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.04794921875,
+      "learning_rate": 0.0001,
+      "loss": 3.0378,
+      "loss/crossentropy": 2.206034767627716,
+      "loss/hidden": 2.803125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24584971666336058,
+      "loss/reg": 0.0,
+      "step": 37550
+    },
+    {
+      "epoch": 0.24710526315789474,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.058288319905598955,
+      "learning_rate": 0.0001,
+      "loss": 2.9179,
+      "loss/crossentropy": 1.987370991706848,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.194708900898695,
+      "loss/reg": 0.0,
+      "step": 37560
+    },
+    {
+      "epoch": 0.24717105263157896,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.03406956990559896,
+      "learning_rate": 0.0001,
+      "loss": 2.9143,
+      "loss/crossentropy": 2.281442165374756,
+      "loss/hidden": 2.69375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19732007533311843,
+      "loss/reg": 0.0,
+      "step": 37570
+    },
+    {
+      "epoch": 0.24723684210526317,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.0267730712890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9615,
+      "loss/crossentropy": 2.3091693341732027,
+      "loss/hidden": 2.6921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22810056209564208,
+      "loss/reg": 0.0,
+      "step": 37580
+    },
+    {
+      "epoch": 0.24730263157894736,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.0711334228515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0095,
+      "loss/crossentropy": 2.276491713523865,
+      "loss/hidden": 2.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2274973288178444,
+      "loss/reg": 0.0,
+      "step": 37590
+    },
+    {
+      "epoch": 0.24736842105263157,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.052994791666666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0237,
+      "loss/crossentropy": 2.096492087841034,
+      "loss/hidden": 2.9484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22436774969100953,
+      "loss/reg": 0.0,
+      "step": 37600
+    },
+    {
+      "epoch": 0.24743421052631578,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.023664347330729165,
+      "learning_rate": 0.0001,
+      "loss": 2.9825,
+      "loss/crossentropy": 2.1003798633813857,
+      "loss/hidden": 2.6734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21021526902914048,
+      "loss/reg": 0.0,
+      "step": 37610
+    },
+    {
+      "epoch": 0.2475,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.12235921223958333,
+      "learning_rate": 0.0001,
+      "loss": 2.9975,
+      "loss/crossentropy": 2.338260293006897,
+      "loss/hidden": 2.6953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2373722493648529,
+      "loss/reg": 0.0,
+      "step": 37620
+    },
+    {
+      "epoch": 0.2475657894736842,
+      "grad_norm": 2.75,
+      "grad_norm_var": 0.1563738505045573,
+      "learning_rate": 0.0001,
+      "loss": 2.9835,
+      "loss/crossentropy": 2.4219950675964355,
+      "loss/hidden": 2.6015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.210829646140337,
+      "loss/reg": 0.0,
+      "step": 37630
+    },
+    {
+      "epoch": 0.24763157894736842,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 3.647915690527556e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1081,
+      "loss/crossentropy": 2.1713775038719176,
+      "loss/hidden": 2.603125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20183515548706055,
+      "loss/reg": 0.0,
+      "step": 37640
+    },
+    {
+      "epoch": 0.24769736842105264,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 3.6479156911331085e+17,
+      "learning_rate": 0.0001,
+      "loss": 2.9457,
+      "loss/crossentropy": 2.509985637664795,
+      "loss/hidden": 2.653125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23514220416545867,
+      "loss/reg": 0.0,
+      "step": 37650
+    },
+    {
+      "epoch": 0.24776315789473685,
+      "grad_norm": 4.125,
+      "grad_norm_var": 0.30654703776041664,
+      "learning_rate": 0.0001,
+      "loss": 2.934,
+      "loss/crossentropy": 2.2876363396644592,
+      "loss/hidden": 2.640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20297650545835494,
+      "loss/reg": 0.0,
+      "step": 37660
+    },
+    {
+      "epoch": 0.24782894736842107,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.27844416300455727,
+      "learning_rate": 0.0001,
+      "loss": 2.9907,
+      "loss/crossentropy": 2.4115204930305483,
+      "loss/hidden": 2.8875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2483261451125145,
+      "loss/reg": 0.0,
+      "step": 37670
+    },
+    {
+      "epoch": 0.24789473684210525,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.05608317057291667,
+      "learning_rate": 0.0001,
+      "loss": 2.9908,
+      "loss/crossentropy": 2.3460915803909304,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24570741802453994,
+      "loss/reg": 0.0,
+      "step": 37680
+    },
+    {
+      "epoch": 0.24796052631578946,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.1174468994140625,
+      "learning_rate": 0.0001,
+      "loss": 2.9498,
+      "loss/crossentropy": 2.2238924860954286,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21785593777894974,
+      "loss/reg": 0.0,
+      "step": 37690
+    },
+    {
+      "epoch": 0.24802631578947368,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.18594462076822918,
+      "learning_rate": 0.0001,
+      "loss": 2.9304,
+      "loss/crossentropy": 2.2358608484268188,
+      "loss/hidden": 2.7125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20752606838941573,
+      "loss/reg": 0.0,
+      "step": 37700
+    },
+    {
+      "epoch": 0.2480921052631579,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 6.753564453125,
+      "learning_rate": 0.0001,
+      "loss": 3.002,
+      "loss/crossentropy": 2.481326103210449,
+      "loss/hidden": 2.96875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.34728084355592725,
+      "loss/reg": 0.0,
+      "step": 37710
+    },
+    {
+      "epoch": 0.2481578947368421,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 6.80933837890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9933,
+      "loss/crossentropy": 2.11043701171875,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.231504824757576,
+      "loss/reg": 0.0,
+      "step": 37720
+    },
+    {
+      "epoch": 0.24822368421052632,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 0.0999176025390625,
+      "learning_rate": 0.0001,
+      "loss": 2.998,
+      "loss/crossentropy": 2.3754311203956604,
+      "loss/hidden": 2.7375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2291927695274353,
+      "loss/reg": 0.0,
+      "step": 37730
+    },
+    {
+      "epoch": 0.24828947368421053,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.044677734375,
+      "learning_rate": 0.0001,
+      "loss": 2.9851,
+      "loss/crossentropy": 2.2369036078453064,
+      "loss/hidden": 2.5625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19738300889730453,
+      "loss/reg": 0.0,
+      "step": 37740
+    },
+    {
+      "epoch": 0.24835526315789475,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.265234375,
+      "learning_rate": 0.0001,
+      "loss": 3.0716,
+      "loss/crossentropy": 2.236533558368683,
+      "loss/hidden": 3.1171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2648838981986046,
+      "loss/reg": 0.0,
+      "step": 37750
+    },
+    {
+      "epoch": 0.24842105263157896,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.28010660807291665,
+      "learning_rate": 0.0001,
+      "loss": 3.0245,
+      "loss/crossentropy": 2.3721311211586,
+      "loss/hidden": 2.790625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27348231226205827,
+      "loss/reg": 0.0,
+      "step": 37760
+    },
+    {
+      "epoch": 0.24848684210526314,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.13202718098958333,
+      "learning_rate": 0.0001,
+      "loss": 3.0549,
+      "loss/crossentropy": 2.2753346085548403,
+      "loss/hidden": 3.0125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2692110911011696,
+      "loss/reg": 0.0,
+      "step": 37770
+    },
+    {
+      "epoch": 0.24855263157894736,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.10732014973958333,
+      "learning_rate": 0.0001,
+      "loss": 3.022,
+      "loss/crossentropy": 2.2350252270698547,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19948195964097976,
+      "loss/reg": 0.0,
+      "step": 37780
+    },
+    {
+      "epoch": 0.24861842105263157,
+      "grad_norm": 3.375,
+      "grad_norm_var": 0.13839518229166667,
+      "learning_rate": 0.0001,
+      "loss": 2.9917,
+      "loss/crossentropy": 2.175236439704895,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21911766976118088,
+      "loss/reg": 0.0,
+      "step": 37790
+    },
+    {
+      "epoch": 0.24868421052631579,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.10273030598958334,
+      "learning_rate": 0.0001,
+      "loss": 2.9491,
+      "loss/crossentropy": 2.3891021251678466,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2441677287220955,
+      "loss/reg": 0.0,
+      "step": 37800
+    },
+    {
+      "epoch": 0.24875,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.11543782552083333,
+      "learning_rate": 0.0001,
+      "loss": 2.9554,
+      "loss/crossentropy": 2.1731114864349363,
+      "loss/hidden": 2.8859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24577540904283524,
+      "loss/reg": 0.0,
+      "step": 37810
+    },
+    {
+      "epoch": 0.2488157894736842,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.13479588826497396,
+      "learning_rate": 0.0001,
+      "loss": 3.0308,
+      "loss/crossentropy": 2.459308052062988,
+      "loss/hidden": 2.8546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25517907589673994,
+      "loss/reg": 0.0,
+      "step": 37820
+    },
+    {
+      "epoch": 0.24888157894736843,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.06494140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0135,
+      "loss/crossentropy": 2.296464502811432,
+      "loss/hidden": 2.8265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21491459012031555,
+      "loss/reg": 0.0,
+      "step": 37830
+    },
+    {
+      "epoch": 0.24894736842105264,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.64693603515625,
+      "learning_rate": 0.0001,
+      "loss": 3.0707,
+      "loss/crossentropy": 2.414829707145691,
+      "loss/hidden": 2.65625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2108585089445114,
+      "loss/reg": 0.0,
+      "step": 37840
+    },
+    {
+      "epoch": 0.24901315789473685,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.5141021728515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9619,
+      "loss/crossentropy": 1.9343510389328002,
+      "loss/hidden": 2.775,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22914507612586021,
+      "loss/reg": 0.0,
+      "step": 37850
+    },
+    {
+      "epoch": 0.24907894736842107,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.1226226806640625,
+      "learning_rate": 0.0001,
+      "loss": 2.967,
+      "loss/crossentropy": 2.5030432462692263,
+      "loss/hidden": 2.6140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23077891543507575,
+      "loss/reg": 0.0,
+      "step": 37860
+    },
+    {
+      "epoch": 0.24914473684210525,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.07437235514322917,
+      "learning_rate": 0.0001,
+      "loss": 3.0293,
+      "loss/crossentropy": 2.3556608080863954,
+      "loss/hidden": 2.725,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2477458581328392,
+      "loss/reg": 0.0,
+      "step": 37870
+    },
+    {
+      "epoch": 0.24921052631578947,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.05681050618489583,
+      "learning_rate": 0.0001,
+      "loss": 2.9916,
+      "loss/crossentropy": 2.2093961358070375,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19773360043764115,
+      "loss/reg": 0.0,
+      "step": 37880
+    },
+    {
+      "epoch": 0.24927631578947368,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.032624308268229166,
+      "learning_rate": 0.0001,
+      "loss": 2.9833,
+      "loss/crossentropy": 2.070064514875412,
+      "loss/hidden": 2.621875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20810991451144217,
+      "loss/reg": 0.0,
+      "step": 37890
+    },
+    {
+      "epoch": 0.2493421052631579,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.3275227864583333,
+      "learning_rate": 0.0001,
+      "loss": 3.087,
+      "loss/crossentropy": 2.4348979711532595,
+      "loss/hidden": 2.7703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1990129753947258,
+      "loss/reg": 0.0,
+      "step": 37900
+    },
+    {
+      "epoch": 0.2494078947368421,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.1884661356608073,
+      "learning_rate": 0.0001,
+      "loss": 2.9224,
+      "loss/crossentropy": 2.313380515575409,
+      "loss/hidden": 2.6796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20934066772460938,
+      "loss/reg": 0.0,
+      "step": 37910
+    },
+    {
+      "epoch": 0.24947368421052632,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.2013567606608073,
+      "learning_rate": 0.0001,
+      "loss": 2.9925,
+      "loss/crossentropy": 2.2767374873161317,
+      "loss/hidden": 2.7171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21055591180920602,
+      "loss/reg": 0.0,
+      "step": 37920
+    },
+    {
+      "epoch": 0.24953947368421053,
+      "grad_norm": 2.921875,
+      "grad_norm_var": 0.29850972493489586,
+      "learning_rate": 0.0001,
+      "loss": 2.9993,
+      "loss/crossentropy": 2.361280381679535,
+      "loss/hidden": 2.75625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2328987866640091,
+      "loss/reg": 0.0,
+      "step": 37930
+    },
+    {
+      "epoch": 0.24960526315789475,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.2740234375,
+      "learning_rate": 0.0001,
+      "loss": 2.9261,
+      "loss/crossentropy": 2.1607108235359194,
+      "loss/hidden": 2.828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21131338626146318,
+      "loss/reg": 0.0,
+      "step": 37940
+    },
+    {
+      "epoch": 0.24967105263157896,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.15468343098958334,
+      "learning_rate": 0.0001,
+      "loss": 2.9839,
+      "loss/crossentropy": 2.2607654333114624,
+      "loss/hidden": 2.6546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18967671394348146,
+      "loss/reg": 0.0,
+      "step": 37950
+    },
+    {
+      "epoch": 0.24973684210526315,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.1874956766764323,
+      "learning_rate": 0.0001,
+      "loss": 3.0293,
+      "loss/crossentropy": 2.483866810798645,
+      "loss/hidden": 2.6421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22582052797079086,
+      "loss/reg": 0.0,
+      "step": 37960
+    },
+    {
+      "epoch": 0.24980263157894736,
+      "grad_norm": 2.40625,
+      "grad_norm_var": 0.21528294881184895,
+      "learning_rate": 0.0001,
+      "loss": 3.017,
+      "loss/crossentropy": 2.568509268760681,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25563120394945144,
+      "loss/reg": 0.0,
+      "step": 37970
+    },
+    {
+      "epoch": 0.24986842105263157,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.21802469889322917,
+      "learning_rate": 0.0001,
+      "loss": 3.002,
+      "loss/crossentropy": 2.394374597072601,
+      "loss/hidden": 2.753125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2696389377117157,
+      "loss/reg": 0.0,
+      "step": 37980
+    },
+    {
+      "epoch": 0.2499342105263158,
+      "grad_norm": 2.296875,
+      "grad_norm_var": 0.41775716145833336,
+      "learning_rate": 0.0001,
+      "loss": 3.0951,
+      "loss/crossentropy": 2.279202771186829,
+      "loss/hidden": 2.853125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22951147556304932,
+      "loss/reg": 0.0,
+      "step": 37990
+    },
+    {
+      "epoch": 0.25,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.25235087076822915,
+      "learning_rate": 0.0001,
+      "loss": 2.9227,
+      "loss/crossentropy": 2.0242787480354307,
+      "loss/hidden": 2.65625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20535071119666098,
+      "loss/reg": 0.0,
+      "step": 38000
+    },
+    {
+      "epoch": 0.2500657894736842,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.08347880045572917,
+      "learning_rate": 0.0001,
+      "loss": 2.9955,
+      "loss/crossentropy": 2.109722208976746,
+      "loss/hidden": 2.83125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20487915128469467,
+      "loss/reg": 0.0,
+      "step": 38010
+    },
+    {
+      "epoch": 0.2501315789473684,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 6.417805380649222e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1302,
+      "loss/crossentropy": 2.31266930103302,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22627270370721816,
+      "loss/reg": 0.0,
+      "step": 38020
+    },
+    {
+      "epoch": 0.2501973684210526,
+      "grad_norm": 2.84375,
+      "grad_norm_var": 0.0590240478515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9995,
+      "loss/crossentropy": 2.147220182418823,
+      "loss/hidden": 2.6078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18802944347262382,
+      "loss/reg": 0.0,
+      "step": 38030
+    },
+    {
+      "epoch": 0.25026315789473685,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.18092041015625,
+      "learning_rate": 0.0001,
+      "loss": 3.014,
+      "loss/crossentropy": 2.1841834664344786,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23548600152134896,
+      "loss/reg": 0.0,
+      "step": 38040
+    },
+    {
+      "epoch": 0.25032894736842104,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.17408854166666668,
+      "learning_rate": 0.0001,
+      "loss": 3.0632,
+      "loss/crossentropy": 2.1282356858253477,
+      "loss/hidden": 2.70625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2339039534330368,
+      "loss/reg": 0.0,
+      "step": 38050
+    },
+    {
+      "epoch": 0.2503947368421053,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.15420633951822918,
+      "learning_rate": 0.0001,
+      "loss": 2.9998,
+      "loss/crossentropy": 2.2876673698425294,
+      "loss/hidden": 2.621875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20793070793151855,
+      "loss/reg": 0.0,
+      "step": 38060
+    },
+    {
+      "epoch": 0.25046052631578947,
+      "grad_norm": 2.765625,
+      "grad_norm_var": 0.17851460774739583,
+      "learning_rate": 0.0001,
+      "loss": 3.0085,
+      "loss/crossentropy": 2.327640974521637,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26281672716140747,
+      "loss/reg": 0.0,
+      "step": 38070
+    },
+    {
+      "epoch": 0.2505263157894737,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.6514231363932291,
+      "learning_rate": 0.0001,
+      "loss": 3.0383,
+      "loss/crossentropy": 1.913532590866089,
+      "loss/hidden": 2.725,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20542303174734117,
+      "loss/reg": 0.0,
+      "step": 38080
+    },
+    {
+      "epoch": 0.2505921052631579,
+      "grad_norm": 1.984375,
+      "grad_norm_var": 0.7180580139160156,
+      "learning_rate": 0.0001,
+      "loss": 2.9836,
+      "loss/crossentropy": 2.27760968208313,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23442062735557556,
+      "loss/reg": 0.0,
+      "step": 38090
+    },
+    {
+      "epoch": 0.2506578947368421,
+      "grad_norm": 4.125,
+      "grad_norm_var": 0.5421974182128906,
+      "learning_rate": 0.0001,
+      "loss": 2.9747,
+      "loss/crossentropy": 2.13877215385437,
+      "loss/hidden": 2.7046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19861070141196252,
+      "loss/reg": 0.0,
+      "step": 38100
+    },
+    {
+      "epoch": 0.2507236842105263,
+      "grad_norm": 3.59375,
+      "grad_norm_var": 0.27887369791666666,
+      "learning_rate": 0.0001,
+      "loss": 3.1014,
+      "loss/crossentropy": 2.2125035911798476,
+      "loss/hidden": 2.7234375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20661384277045727,
+      "loss/reg": 0.0,
+      "step": 38110
+    },
+    {
+      "epoch": 0.2507894736842105,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.3249908447265625,
+      "learning_rate": 0.0001,
+      "loss": 3.0608,
+      "loss/crossentropy": 2.3948875069618225,
+      "loss/hidden": 2.671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21930506974458694,
+      "loss/reg": 0.0,
+      "step": 38120
+    },
+    {
+      "epoch": 0.25085526315789475,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.28649063110351564,
+      "learning_rate": 0.0001,
+      "loss": 2.9885,
+      "loss/crossentropy": 2.292639744281769,
+      "loss/hidden": 2.7921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2267945870757103,
+      "loss/reg": 0.0,
+      "step": 38130
+    },
+    {
+      "epoch": 0.25092105263157893,
+      "grad_norm": 2.59375,
+      "grad_norm_var": 0.05728759765625,
+      "learning_rate": 0.0001,
+      "loss": 2.9936,
+      "loss/crossentropy": 2.097894775867462,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20571077913045882,
+      "loss/reg": 0.0,
+      "step": 38140
+    },
+    {
+      "epoch": 0.2509868421052632,
+      "grad_norm": 3.375,
+      "grad_norm_var": 3.1128214518229167,
+      "learning_rate": 0.0001,
+      "loss": 2.9914,
+      "loss/crossentropy": 2.272007715702057,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21056424751877784,
+      "loss/reg": 0.0,
+      "step": 38150
+    },
+    {
+      "epoch": 0.25105263157894736,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.16922098795572918,
+      "learning_rate": 0.0001,
+      "loss": 2.9551,
+      "loss/crossentropy": 2.2579661190509794,
+      "loss/hidden": 2.753125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26000856757164004,
+      "loss/reg": 0.0,
+      "step": 38160
+    },
+    {
+      "epoch": 0.2511184210526316,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.1134674072265625,
+      "learning_rate": 0.0001,
+      "loss": 2.9975,
+      "loss/crossentropy": 2.0489343285560606,
+      "loss/hidden": 2.6265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.17201969623565674,
+      "loss/reg": 0.0,
+      "step": 38170
+    },
+    {
+      "epoch": 0.2511842105263158,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.030980428059895832,
+      "learning_rate": 0.0001,
+      "loss": 2.9523,
+      "loss/crossentropy": 2.252878558635712,
+      "loss/hidden": 2.8890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23066311478614807,
+      "loss/reg": 0.0,
+      "step": 38180
+    },
+    {
+      "epoch": 0.25125,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.0248931884765625,
+      "learning_rate": 0.0001,
+      "loss": 2.9488,
+      "loss/crossentropy": 2.3150659799575806,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21407101899385453,
+      "loss/reg": 0.0,
+      "step": 38190
+    },
+    {
+      "epoch": 0.2513157894736842,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 2.1658424377441405,
+      "learning_rate": 0.0001,
+      "loss": 2.9789,
+      "loss/crossentropy": 2.4279683232307434,
+      "loss/hidden": 2.696875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2308952897787094,
+      "loss/reg": 0.0,
+      "step": 38200
+    },
+    {
+      "epoch": 0.2513815789473684,
+      "grad_norm": 3.75,
+      "grad_norm_var": 1.915710194905599,
+      "learning_rate": 0.0001,
+      "loss": 2.9791,
+      "loss/crossentropy": 2.454149055480957,
+      "loss/hidden": 2.778125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27904040217399595,
+      "loss/reg": 0.0,
+      "step": 38210
+    },
+    {
+      "epoch": 0.25144736842105264,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.16384175618489583,
+      "learning_rate": 0.0001,
+      "loss": 3.025,
+      "loss/crossentropy": 2.4395872354507446,
+      "loss/hidden": 2.7875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24970148950815202,
+      "loss/reg": 0.0,
+      "step": 38220
+    },
+    {
+      "epoch": 0.2515131578947368,
+      "grad_norm": 3.1875,
+      "grad_norm_var": 0.13736572265625,
+      "learning_rate": 0.0001,
+      "loss": 2.9912,
+      "loss/crossentropy": 2.4788294553756716,
+      "loss/hidden": 2.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21640673130750657,
+      "loss/reg": 0.0,
+      "step": 38230
+    },
+    {
+      "epoch": 0.25157894736842107,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.1446929931640625,
+      "learning_rate": 0.0001,
+      "loss": 2.978,
+      "loss/crossentropy": 2.15856648683548,
+      "loss/hidden": 2.9171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22779612690210344,
+      "loss/reg": 0.0,
+      "step": 38240
+    },
+    {
+      "epoch": 0.25164473684210525,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.028629557291666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0221,
+      "loss/crossentropy": 2.5230648517608643,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22997782975435258,
+      "loss/reg": 0.0,
+      "step": 38250
+    },
+    {
+      "epoch": 0.2517105263157895,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.03369852701822917,
+      "learning_rate": 0.0001,
+      "loss": 2.9928,
+      "loss/crossentropy": 2.3802834033966063,
+      "loss/hidden": 2.8765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25779378563165667,
+      "loss/reg": 0.0,
+      "step": 38260
+    },
+    {
+      "epoch": 0.2517763157894737,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.18492838541666667,
+      "learning_rate": 0.0001,
+      "loss": 2.9498,
+      "loss/crossentropy": 2.31593804359436,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.245054192841053,
+      "loss/reg": 0.0,
+      "step": 38270
+    },
+    {
+      "epoch": 0.25184210526315787,
+      "grad_norm": 3.0625,
+      "grad_norm_var": 0.1485015869140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0313,
+      "loss/crossentropy": 2.241119909286499,
+      "loss/hidden": 2.60625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19746284037828446,
+      "loss/reg": 0.0,
+      "step": 38280
+    },
+    {
+      "epoch": 0.2519078947368421,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.10631510416666666,
+      "learning_rate": 0.0001,
+      "loss": 3.0635,
+      "loss/crossentropy": 2.0983599185943604,
+      "loss/hidden": 2.99375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25142730176448824,
+      "loss/reg": 0.0,
+      "step": 38290
+    },
+    {
+      "epoch": 0.2519736842105263,
+      "grad_norm": 3.28125,
+      "grad_norm_var": 0.11451416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.046,
+      "loss/crossentropy": 2.36417875289917,
+      "loss/hidden": 2.7890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24252529442310333,
+      "loss/reg": 0.0,
+      "step": 38300
+    },
+    {
+      "epoch": 0.25203947368421054,
+      "grad_norm": 3.046875,
+      "grad_norm_var": 0.11712239583333334,
+      "learning_rate": 0.0001,
+      "loss": 3.1178,
+      "loss/crossentropy": 2.3146223425865173,
+      "loss/hidden": 2.71875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26736019253730775,
+      "loss/reg": 0.0,
+      "step": 38310
+    },
+    {
+      "epoch": 0.2521052631578947,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.7847076416015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0818,
+      "loss/crossentropy": 2.2367316365242003,
+      "loss/hidden": 2.5859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19362774714827538,
+      "loss/reg": 0.0,
+      "step": 38320
+    },
+    {
+      "epoch": 0.25217105263157896,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.3490386962890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9877,
+      "loss/crossentropy": 2.212986183166504,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21934635937213898,
+      "loss/reg": 0.0,
+      "step": 38330
+    },
+    {
+      "epoch": 0.25223684210526315,
+      "grad_norm": 2.71875,
+      "grad_norm_var": 0.0618560791015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0083,
+      "loss/crossentropy": 2.424324858188629,
+      "loss/hidden": 2.6875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2144932597875595,
+      "loss/reg": 0.0,
+      "step": 38340
+    },
+    {
+      "epoch": 0.2523026315789474,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.05483296712239583,
+      "learning_rate": 0.0001,
+      "loss": 2.9994,
+      "loss/crossentropy": 2.3819736361503603,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.282964862883091,
+      "loss/reg": 0.0,
+      "step": 38350
+    },
+    {
+      "epoch": 0.2523684210526316,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 6.966681585372321e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0863,
+      "loss/crossentropy": 2.094909679889679,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21131259053945542,
+      "loss/reg": 0.0,
+      "step": 38360
+    },
+    {
+      "epoch": 0.2524342105263158,
+      "grad_norm": 1.9765625,
+      "grad_norm_var": 0.49862848917643227,
+      "learning_rate": 0.0001,
+      "loss": 2.9827,
+      "loss/crossentropy": 2.215614175796509,
+      "loss/hidden": 2.7109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2258283868432045,
+      "loss/reg": 0.0,
+      "step": 38370
+    },
+    {
+      "epoch": 0.2525,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.4853370666503906,
+      "learning_rate": 0.0001,
+      "loss": 2.9659,
+      "loss/crossentropy": 2.0444746017456055,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.17757057920098304,
+      "loss/reg": 0.0,
+      "step": 38380
+    },
+    {
+      "epoch": 0.2525657894736842,
+      "grad_norm": 2.4375,
+      "grad_norm_var": 0.09589436848958334,
+      "learning_rate": 0.0001,
+      "loss": 3.0212,
+      "loss/crossentropy": 2.3424991607666015,
+      "loss/hidden": 2.6484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20197789669036864,
+      "loss/reg": 0.0,
+      "step": 38390
+    },
+    {
+      "epoch": 0.25263157894736843,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.10813395182291667,
+      "learning_rate": 0.0001,
+      "loss": 2.9537,
+      "loss/crossentropy": 2.090820240974426,
+      "loss/hidden": 2.6109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18726751990616322,
+      "loss/reg": 0.0,
+      "step": 38400
+    },
+    {
+      "epoch": 0.2526973684210526,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.06373291015625,
+      "learning_rate": 0.0001,
+      "loss": 2.9561,
+      "loss/crossentropy": 1.846102112531662,
+      "loss/hidden": 2.653125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1744068369269371,
+      "loss/reg": 0.0,
+      "step": 38410
+    },
+    {
+      "epoch": 0.25276315789473686,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.10798924763997396,
+      "learning_rate": 0.0001,
+      "loss": 2.9147,
+      "loss/crossentropy": 2.336298942565918,
+      "loss/hidden": 2.753125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2202483594417572,
+      "loss/reg": 0.0,
+      "step": 38420
+    },
+    {
+      "epoch": 0.25282894736842104,
+      "grad_norm": 2.671875,
+      "grad_norm_var": 0.21279195149739583,
+      "learning_rate": 0.0001,
+      "loss": 3.0227,
+      "loss/crossentropy": 2.413504159450531,
+      "loss/hidden": 2.8359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2817773073911667,
+      "loss/reg": 0.0,
+      "step": 38430
+    },
+    {
+      "epoch": 0.2528947368421053,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.4711985270182292,
+      "learning_rate": 0.0001,
+      "loss": 2.9499,
+      "loss/crossentropy": 2.3754114389419554,
+      "loss/hidden": 2.6328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20899627953767777,
+      "loss/reg": 0.0,
+      "step": 38440
+    },
+    {
+      "epoch": 0.25296052631578947,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.11868489583333333,
+      "learning_rate": 0.0001,
+      "loss": 3.0609,
+      "loss/crossentropy": 2.1895971417427065,
+      "loss/hidden": 2.9484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27013812959194183,
+      "loss/reg": 0.0,
+      "step": 38450
+    },
+    {
+      "epoch": 0.2530263157894737,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.07203776041666667,
+      "learning_rate": 0.0001,
+      "loss": 2.9742,
+      "loss/crossentropy": 2.398327910900116,
+      "loss/hidden": 2.534375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20559832453727722,
+      "loss/reg": 0.0,
+      "step": 38460
+    },
+    {
+      "epoch": 0.2530921052631579,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.042479451497395834,
+      "learning_rate": 0.0001,
+      "loss": 3.0595,
+      "loss/crossentropy": 2.1632952094078064,
+      "loss/hidden": 3.0921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26632697582244874,
+      "loss/reg": 0.0,
+      "step": 38470
+    },
+    {
+      "epoch": 0.2531578947368421,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.0286285400390625,
+      "learning_rate": 0.0001,
+      "loss": 2.948,
+      "loss/crossentropy": 2.3557206630706786,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23956954926252366,
+      "loss/reg": 0.0,
+      "step": 38480
+    },
+    {
+      "epoch": 0.2532236842105263,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.06353759765625,
+      "learning_rate": 0.0001,
+      "loss": 3.0005,
+      "loss/crossentropy": 2.1524956703186033,
+      "loss/hidden": 2.6375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2026323951780796,
+      "loss/reg": 0.0,
+      "step": 38490
+    },
+    {
+      "epoch": 0.2532894736842105,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.04701309204101563,
+      "learning_rate": 0.0001,
+      "loss": 2.9439,
+      "loss/crossentropy": 2.081376886367798,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20958617925643921,
+      "loss/reg": 0.0,
+      "step": 38500
+    },
+    {
+      "epoch": 0.25335526315789475,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.057708485921223955,
+      "learning_rate": 0.0001,
+      "loss": 3.008,
+      "loss/crossentropy": 2.419314205646515,
+      "loss/hidden": 2.9125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.28001276552677157,
+      "loss/reg": 0.0,
+      "step": 38510
+    },
+    {
+      "epoch": 0.25342105263157894,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 280.5541015625,
+      "learning_rate": 0.0001,
+      "loss": 3.124,
+      "loss/crossentropy": 2.209311616420746,
+      "loss/hidden": 3.334375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.4400923550128937,
+      "loss/reg": 0.0,
+      "step": 38520
+    },
+    {
+      "epoch": 0.2534868421052632,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 280.8059855143229,
+      "learning_rate": 0.0001,
+      "loss": 2.963,
+      "loss/crossentropy": 2.386173665523529,
+      "loss/hidden": 2.6765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20247507095336914,
+      "loss/reg": 0.0,
+      "step": 38530
+    },
+    {
+      "epoch": 0.25355263157894736,
+      "grad_norm": 2.375,
+      "grad_norm_var": 1.2670237223307292,
+      "learning_rate": 0.0001,
+      "loss": 2.9903,
+      "loss/crossentropy": 2.3542658567428587,
+      "loss/hidden": 2.759375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.229895980656147,
+      "loss/reg": 0.0,
+      "step": 38540
+    },
+    {
+      "epoch": 0.2536184210526316,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.12191340128580729,
+      "learning_rate": 0.0001,
+      "loss": 3.0164,
+      "loss/crossentropy": 2.4656338930130004,
+      "loss/hidden": 2.653125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2140819951891899,
+      "loss/reg": 0.0,
+      "step": 38550
+    },
+    {
+      "epoch": 0.2536842105263158,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 9.503270467122396,
+      "learning_rate": 0.0001,
+      "loss": 3.0177,
+      "loss/crossentropy": 2.4247732520103455,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26161234080791473,
+      "loss/reg": 0.0,
+      "step": 38560
+    },
+    {
+      "epoch": 0.25375,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 9.574365234375,
+      "learning_rate": 0.0001,
+      "loss": 3.0298,
+      "loss/crossentropy": 2.2777958273887635,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.3407336473464966,
+      "loss/reg": 0.0,
+      "step": 38570
+    },
+    {
+      "epoch": 0.2538157894736842,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 0.3904693603515625,
+      "learning_rate": 0.0001,
+      "loss": 2.933,
+      "loss/crossentropy": 2.220312762260437,
+      "loss/hidden": 2.7015625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2032586969435215,
+      "loss/reg": 0.0,
+      "step": 38580
+    },
+    {
+      "epoch": 0.2538815789473684,
+      "grad_norm": 2.78125,
+      "grad_norm_var": 0.08896077473958333,
+      "learning_rate": 0.0001,
+      "loss": 2.9606,
+      "loss/crossentropy": 2.1711514949798585,
+      "loss/hidden": 2.646875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.17734147608280182,
+      "loss/reg": 0.0,
+      "step": 38590
+    },
+    {
+      "epoch": 0.25394736842105264,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.20473531087239583,
+      "learning_rate": 0.0001,
+      "loss": 2.9321,
+      "loss/crossentropy": 1.9470943689346314,
+      "loss/hidden": 2.6,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19334470629692077,
+      "loss/reg": 0.0,
+      "step": 38600
+    },
+    {
+      "epoch": 0.25401315789473683,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.06678059895833334,
+      "learning_rate": 0.0001,
+      "loss": 2.9233,
+      "loss/crossentropy": 2.273166114091873,
+      "loss/hidden": 2.7109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20435749739408493,
+      "loss/reg": 0.0,
+      "step": 38610
+    },
+    {
+      "epoch": 0.25407894736842107,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.11378580729166667,
+      "learning_rate": 0.0001,
+      "loss": 2.9767,
+      "loss/crossentropy": 2.571880316734314,
+      "loss/hidden": 2.6453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21721504628658295,
+      "loss/reg": 0.0,
+      "step": 38620
+    },
+    {
+      "epoch": 0.25414473684210526,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.08742574055989584,
+      "learning_rate": 0.0001,
+      "loss": 2.9226,
+      "loss/crossentropy": 2.2264232873916625,
+      "loss/hidden": 2.890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21014079004526137,
+      "loss/reg": 0.0,
+      "step": 38630
+    },
+    {
+      "epoch": 0.2542105263157895,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.05276285807291667,
+      "learning_rate": 0.0001,
+      "loss": 2.9634,
+      "loss/crossentropy": 2.2364309787750245,
+      "loss/hidden": 2.6953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20938622057437897,
+      "loss/reg": 0.0,
+      "step": 38640
+    },
+    {
+      "epoch": 0.2542763157894737,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.03668619791666667,
+      "learning_rate": 0.0001,
+      "loss": 3.0195,
+      "loss/crossentropy": 2.135267126560211,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2628340318799019,
+      "loss/reg": 0.0,
+      "step": 38650
+    },
+    {
+      "epoch": 0.25434210526315787,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 1.3797353108723958,
+      "learning_rate": 0.0001,
+      "loss": 2.927,
+      "loss/crossentropy": 2.403776562213898,
+      "loss/hidden": 2.6421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2078718587756157,
+      "loss/reg": 0.0,
+      "step": 38660
+    },
+    {
+      "epoch": 0.2544078947368421,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 1.3981770833333333,
+      "learning_rate": 0.0001,
+      "loss": 2.9993,
+      "loss/crossentropy": 2.358154129981995,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2573642887175083,
+      "loss/reg": 0.0,
+      "step": 38670
+    },
+    {
+      "epoch": 0.2544736842105263,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.06347249348958334,
+      "learning_rate": 0.0001,
+      "loss": 2.9781,
+      "loss/crossentropy": 2.3121828198432923,
+      "loss/hidden": 2.6828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2003457449376583,
+      "loss/reg": 0.0,
+      "step": 38680
+    },
+    {
+      "epoch": 0.25453947368421054,
+      "grad_norm": 6.4375,
+      "grad_norm_var": 1.1110694885253907,
+      "learning_rate": 0.0001,
+      "loss": 2.9505,
+      "loss/crossentropy": 2.153343600034714,
+      "loss/hidden": 2.9796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25137421637773516,
+      "loss/reg": 0.0,
+      "step": 38690
+    },
+    {
+      "epoch": 0.2546052631578947,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 1.0830393473307292,
+      "learning_rate": 0.0001,
+      "loss": 2.9718,
+      "loss/crossentropy": 2.052643448114395,
+      "loss/hidden": 2.9171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25727019011974334,
+      "loss/reg": 0.0,
+      "step": 38700
+    },
+    {
+      "epoch": 0.25467105263157896,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.0537506103515625,
+      "learning_rate": 0.0001,
+      "loss": 3.006,
+      "loss/crossentropy": 2.152147728204727,
+      "loss/hidden": 2.8328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23068247735500336,
+      "loss/reg": 0.0,
+      "step": 38710
+    },
+    {
+      "epoch": 0.25473684210526315,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.0999420166015625,
+      "learning_rate": 0.0001,
+      "loss": 2.9765,
+      "loss/crossentropy": 2.2647915482521057,
+      "loss/hidden": 2.678125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24034264236688613,
+      "loss/reg": 0.0,
+      "step": 38720
+    },
+    {
+      "epoch": 0.2548026315789474,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.19918212890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9698,
+      "loss/crossentropy": 2.4958234548568727,
+      "loss/hidden": 2.7671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23010408282279968,
+      "loss/reg": 0.0,
+      "step": 38730
+    },
+    {
+      "epoch": 0.2548684210526316,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.15207926432291666,
+      "learning_rate": 0.0001,
+      "loss": 2.9802,
+      "loss/crossentropy": 2.2727454662323,
+      "loss/hidden": 2.696875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20178039968013764,
+      "loss/reg": 0.0,
+      "step": 38740
+    },
+    {
+      "epoch": 0.25493421052631576,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.06728108723958333,
+      "learning_rate": 0.0001,
+      "loss": 3.036,
+      "loss/crossentropy": 2.2367177844047545,
+      "loss/hidden": 2.865625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24128163158893584,
+      "loss/reg": 0.0,
+      "step": 38750
+    },
+    {
+      "epoch": 0.255,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.06168212890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9621,
+      "loss/crossentropy": 2.34770849943161,
+      "loss/hidden": 2.7171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21328310072422027,
+      "loss/reg": 0.0,
+      "step": 38760
+    },
+    {
+      "epoch": 0.2550657894736842,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.1226470947265625,
+      "learning_rate": 0.0001,
+      "loss": 2.9942,
+      "loss/crossentropy": 2.306888747215271,
+      "loss/hidden": 2.8390625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2507788211107254,
+      "loss/reg": 0.0,
+      "step": 38770
+    },
+    {
+      "epoch": 0.25513157894736843,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.17222900390625,
+      "learning_rate": 0.0001,
+      "loss": 2.8878,
+      "loss/crossentropy": 2.0050193548202513,
+      "loss/hidden": 2.7515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18458792492747306,
+      "loss/reg": 0.0,
+      "step": 38780
+    },
+    {
+      "epoch": 0.2551973684210526,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.0305816650390625,
+      "learning_rate": 0.0001,
+      "loss": 2.9807,
+      "loss/crossentropy": 2.279655563831329,
+      "loss/hidden": 2.7359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22222616970539094,
+      "loss/reg": 0.0,
+      "step": 38790
+    },
+    {
+      "epoch": 0.25526315789473686,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.045426432291666666,
+      "learning_rate": 0.0001,
+      "loss": 2.9266,
+      "loss/crossentropy": 1.8874349355697633,
+      "loss/hidden": 2.6609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1566801816225052,
+      "loss/reg": 0.0,
+      "step": 38800
+    },
+    {
+      "epoch": 0.25532894736842104,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.13137919108072918,
+      "learning_rate": 0.0001,
+      "loss": 2.9532,
+      "loss/crossentropy": 2.0287832260131835,
+      "loss/hidden": 2.8375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19014589190483094,
+      "loss/reg": 0.0,
+      "step": 38810
+    },
+    {
+      "epoch": 0.2553947368421053,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.16416727701822917,
+      "learning_rate": 0.0001,
+      "loss": 2.9798,
+      "loss/crossentropy": 2.183877873420715,
+      "loss/hidden": 2.690625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21431390047073365,
+      "loss/reg": 0.0,
+      "step": 38820
+    },
+    {
+      "epoch": 0.25546052631578947,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.24260965983072916,
+      "learning_rate": 0.0001,
+      "loss": 2.9393,
+      "loss/crossentropy": 2.440583086013794,
+      "loss/hidden": 2.6515625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20157338082790374,
+      "loss/reg": 0.0,
+      "step": 38830
+    },
+    {
+      "epoch": 0.25552631578947366,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.1785797119140625,
+      "learning_rate": 0.0001,
+      "loss": 2.9731,
+      "loss/crossentropy": 2.5752979278564454,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24179309010505676,
+      "loss/reg": 0.0,
+      "step": 38840
+    },
+    {
+      "epoch": 0.2555921052631579,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.1367340087890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9657,
+      "loss/crossentropy": 2.35681791305542,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24579865038394927,
+      "loss/reg": 0.0,
+      "step": 38850
+    },
+    {
+      "epoch": 0.2556578947368421,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.13662007649739583,
+      "learning_rate": 0.0001,
+      "loss": 2.939,
+      "loss/crossentropy": 2.2475174188613893,
+      "loss/hidden": 2.7375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21202882528305053,
+      "loss/reg": 0.0,
+      "step": 38860
+    },
+    {
+      "epoch": 0.2557236842105263,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.2050933837890625,
+      "learning_rate": 0.0001,
+      "loss": 2.9837,
+      "loss/crossentropy": 2.3431418418884276,
+      "loss/hidden": 2.7859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23483068943023683,
+      "loss/reg": 0.0,
+      "step": 38870
+    },
+    {
+      "epoch": 0.2557894736842105,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 2.675121053059896,
+      "learning_rate": 0.0001,
+      "loss": 2.9856,
+      "loss/crossentropy": 2.483526587486267,
+      "loss/hidden": 2.7828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2897988960146904,
+      "loss/reg": 0.0,
+      "step": 38880
+    },
+    {
+      "epoch": 0.25585526315789475,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.0619781494140625,
+      "learning_rate": 0.0001,
+      "loss": 2.9633,
+      "loss/crossentropy": 2.1767768919467927,
+      "loss/hidden": 2.75,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.234356290102005,
+      "loss/reg": 0.0,
+      "step": 38890
+    },
+    {
+      "epoch": 0.25592105263157894,
+      "grad_norm": 2.96875,
+      "grad_norm_var": 0.2576568603515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9781,
+      "loss/crossentropy": 2.4225967705249785,
+      "loss/hidden": 2.5765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23021362945437432,
+      "loss/reg": 0.0,
+      "step": 38900
+    },
+    {
+      "epoch": 0.2559868421052632,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.21616923014322917,
+      "learning_rate": 0.0001,
+      "loss": 2.9813,
+      "loss/crossentropy": 2.215132641792297,
+      "loss/hidden": 2.74375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23095793277025223,
+      "loss/reg": 0.0,
+      "step": 38910
+    },
+    {
+      "epoch": 0.25605263157894737,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.08826497395833334,
+      "learning_rate": 0.0001,
+      "loss": 2.9896,
+      "loss/crossentropy": 2.1709822535514833,
+      "loss/hidden": 2.7578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20210467725992204,
+      "loss/reg": 0.0,
+      "step": 38920
+    },
+    {
+      "epoch": 0.2561184210526316,
+      "grad_norm": 2.375,
+      "grad_norm_var": 0.08967997233072916,
+      "learning_rate": 0.0001,
+      "loss": 2.9936,
+      "loss/crossentropy": 2.343753528594971,
+      "loss/hidden": 2.8421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22285044342279434,
+      "loss/reg": 0.0,
+      "step": 38930
+    },
+    {
+      "epoch": 0.2561842105263158,
+      "grad_norm": 2.421875,
+      "grad_norm_var": 0.029637654622395832,
+      "learning_rate": 0.0001,
+      "loss": 2.918,
+      "loss/crossentropy": 2.4347527027130127,
+      "loss/hidden": 2.7078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2453876256942749,
+      "loss/reg": 0.0,
+      "step": 38940
+    },
+    {
+      "epoch": 0.25625,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.058405558268229164,
+      "learning_rate": 0.0001,
+      "loss": 2.9806,
+      "loss/crossentropy": 2.24286550283432,
+      "loss/hidden": 2.878125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24712017327547073,
+      "loss/reg": 0.0,
+      "step": 38950
+    },
+    {
+      "epoch": 0.2563157894736842,
+      "grad_norm": 2.390625,
+      "grad_norm_var": 0.051756795247395834,
+      "learning_rate": 0.0001,
+      "loss": 2.9834,
+      "loss/crossentropy": 2.170624256134033,
+      "loss/hidden": 2.7609375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2218889057636261,
+      "loss/reg": 0.0,
+      "step": 38960
+    },
+    {
+      "epoch": 0.2563815789473684,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.0666015625,
+      "learning_rate": 0.0001,
+      "loss": 3.0572,
+      "loss/crossentropy": 2.154159963130951,
+      "loss/hidden": 2.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20279831290245057,
+      "loss/reg": 0.0,
+      "step": 38970
+    },
+    {
+      "epoch": 0.25644736842105265,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.08325093587239583,
+      "learning_rate": 0.0001,
+      "loss": 2.9366,
+      "loss/crossentropy": 2.1254536151885985,
+      "loss/hidden": 2.9421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24286270141601562,
+      "loss/reg": 0.0,
+      "step": 38980
+    },
+    {
+      "epoch": 0.25651315789473683,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.1099609375,
+      "learning_rate": 0.0001,
+      "loss": 2.9717,
+      "loss/crossentropy": 2.1491544008255006,
+      "loss/hidden": 2.915625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27067394405603407,
+      "loss/reg": 0.0,
+      "step": 38990
+    },
+    {
+      "epoch": 0.2565789473684211,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.6976226806640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0136,
+      "loss/crossentropy": 2.2009241580963135,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19883604645729064,
+      "loss/reg": 0.0,
+      "step": 39000
+    },
+    {
+      "epoch": 0.25664473684210526,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.6546183268229167,
+      "learning_rate": 0.0001,
+      "loss": 2.9775,
+      "loss/crossentropy": 2.086053115129471,
+      "loss/hidden": 2.89375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26883399933576585,
+      "loss/reg": 0.0,
+      "step": 39010
+    },
+    {
+      "epoch": 0.2567105263157895,
+      "grad_norm": 3.71875,
+      "grad_norm_var": 0.23528238932291667,
+      "learning_rate": 0.0001,
+      "loss": 2.9801,
+      "loss/crossentropy": 2.121158719062805,
+      "loss/hidden": 2.8046875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22595785260200502,
+      "loss/reg": 0.0,
+      "step": 39020
+    },
+    {
+      "epoch": 0.2567763157894737,
+      "grad_norm": 2.203125,
+      "grad_norm_var": 0.22114156087239584,
+      "learning_rate": 0.0001,
+      "loss": 2.8994,
+      "loss/crossentropy": 2.2681718945503233,
+      "loss/hidden": 2.6765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25748113095760344,
+      "loss/reg": 0.0,
+      "step": 39030
+    },
+    {
+      "epoch": 0.25684210526315787,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.08943583170572916,
+      "learning_rate": 0.0001,
+      "loss": 2.8768,
+      "loss/crossentropy": 2.4575144767761232,
+      "loss/hidden": 2.578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20152714401483535,
+      "loss/reg": 0.0,
+      "step": 39040
+    },
+    {
+      "epoch": 0.2569078947368421,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.0328765869140625,
+      "learning_rate": 0.0001,
+      "loss": 2.9306,
+      "loss/crossentropy": 2.369240176677704,
+      "loss/hidden": 2.809375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22774229645729066,
+      "loss/reg": 0.0,
+      "step": 39050
+    },
+    {
+      "epoch": 0.2569736842105263,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 3.1118817818339136e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0462,
+      "loss/crossentropy": 2.1120986580848693,
+      "loss/hidden": 2.884375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2498351290822029,
+      "loss/reg": 0.0,
+      "step": 39060
+    },
+    {
+      "epoch": 0.25703947368421054,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.034886678059895836,
+      "learning_rate": 0.0001,
+      "loss": 2.8904,
+      "loss/crossentropy": 2.1073123455047607,
+      "loss/hidden": 2.5921875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19584316536784172,
+      "loss/reg": 0.0,
+      "step": 39070
+    },
+    {
+      "epoch": 0.2571052631578947,
+      "grad_norm": 2.890625,
+      "grad_norm_var": 0.07847900390625,
+      "learning_rate": 0.0001,
+      "loss": 2.9755,
+      "loss/crossentropy": 2.3284847140312195,
+      "loss/hidden": 2.8578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2493668183684349,
+      "loss/reg": 0.0,
+      "step": 39080
+    },
+    {
+      "epoch": 0.25717105263157897,
+      "grad_norm": 2.0625,
+      "grad_norm_var": 0.12560221354166667,
+      "learning_rate": 0.0001,
+      "loss": 2.9003,
+      "loss/crossentropy": 2.334082317352295,
+      "loss/hidden": 2.546875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1966678135097027,
+      "loss/reg": 0.0,
+      "step": 39090
+    },
+    {
+      "epoch": 0.25723684210526315,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.08367513020833334,
+      "learning_rate": 0.0001,
+      "loss": 2.9616,
+      "loss/crossentropy": 2.6055662870407104,
+      "loss/hidden": 2.6140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2218620851635933,
+      "loss/reg": 0.0,
+      "step": 39100
+    },
+    {
+      "epoch": 0.2573026315789474,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.22674051920572916,
+      "learning_rate": 0.0001,
+      "loss": 2.9551,
+      "loss/crossentropy": 2.1810374021530152,
+      "loss/hidden": 2.5875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20281792134046556,
+      "loss/reg": 0.0,
+      "step": 39110
+    },
+    {
+      "epoch": 0.2573684210526316,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.13672587076822917,
+      "learning_rate": 0.0001,
+      "loss": 2.9471,
+      "loss/crossentropy": 2.150469708442688,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21119417250156403,
+      "loss/reg": 0.0,
+      "step": 39120
+    },
+    {
+      "epoch": 0.25743421052631577,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.7312978108723959,
+      "learning_rate": 0.0001,
+      "loss": 2.9604,
+      "loss/crossentropy": 1.9471443891525269,
+      "loss/hidden": 2.8109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2092149168252945,
+      "loss/reg": 0.0,
+      "step": 39130
+    },
+    {
+      "epoch": 0.2575,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.7445330301920573,
+      "learning_rate": 0.0001,
+      "loss": 2.9095,
+      "loss/crossentropy": 2.2832953572273254,
+      "loss/hidden": 2.64375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21912914961576463,
+      "loss/reg": 0.0,
+      "step": 39140
+    },
+    {
+      "epoch": 0.2575657894736842,
+      "grad_norm": 3.15625,
+      "grad_norm_var": 0.14458719889322916,
+      "learning_rate": 0.0001,
+      "loss": 2.9719,
+      "loss/crossentropy": 2.3353841066360475,
+      "loss/hidden": 2.6859375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22544662952423095,
+      "loss/reg": 0.0,
+      "step": 39150
+    },
+    {
+      "epoch": 0.25763157894736843,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.3525136311848958,
+      "learning_rate": 0.0001,
+      "loss": 3.0065,
+      "loss/crossentropy": 2.3361639380455017,
+      "loss/hidden": 2.7296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2251232832670212,
+      "loss/reg": 0.0,
+      "step": 39160
+    },
+    {
+      "epoch": 0.2576973684210526,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.295556640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0384,
+      "loss/crossentropy": 2.324616348743439,
+      "loss/hidden": 2.746875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23191166073083877,
+      "loss/reg": 0.0,
+      "step": 39170
+    },
+    {
+      "epoch": 0.25776315789473686,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.06609700520833334,
+      "learning_rate": 0.0001,
+      "loss": 2.9448,
+      "loss/crossentropy": 2.2094304859638214,
+      "loss/hidden": 2.7359375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24130020663142204,
+      "loss/reg": 0.0,
+      "step": 39180
+    },
+    {
+      "epoch": 0.25782894736842105,
+      "grad_norm": 2.03125,
+      "grad_norm_var": 0.09587376912434896,
+      "learning_rate": 0.0001,
+      "loss": 2.9087,
+      "loss/crossentropy": 2.379458463191986,
+      "loss/hidden": 2.70625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2207609809935093,
+      "loss/reg": 0.0,
+      "step": 39190
+    },
+    {
+      "epoch": 0.2578947368421053,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.10582275390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0122,
+      "loss/crossentropy": 2.1732101857662203,
+      "loss/hidden": 2.6109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20026833638548852,
+      "loss/reg": 0.0,
+      "step": 39200
+    },
+    {
+      "epoch": 0.2579605263157895,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.07842992146809896,
+      "learning_rate": 0.0001,
+      "loss": 2.9131,
+      "loss/crossentropy": 1.8863205194473267,
+      "loss/hidden": 2.6109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1805630184710026,
+      "loss/reg": 0.0,
+      "step": 39210
+    },
+    {
+      "epoch": 0.25802631578947366,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.16445719401041667,
+      "learning_rate": 0.0001,
+      "loss": 2.9865,
+      "loss/crossentropy": 2.3708848357200623,
+      "loss/hidden": 2.7453125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2134988009929657,
+      "loss/reg": 0.0,
+      "step": 39220
+    },
+    {
+      "epoch": 0.2580921052631579,
+      "grad_norm": 2.109375,
+      "grad_norm_var": 0.043190256754557295,
+      "learning_rate": 0.0001,
+      "loss": 2.8704,
+      "loss/crossentropy": 2.5470376133918764,
+      "loss/hidden": 2.6828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21798246651887893,
+      "loss/reg": 0.0,
+      "step": 39230
+    },
+    {
+      "epoch": 0.2581578947368421,
+      "grad_norm": 3.75,
+      "grad_norm_var": 0.15417378743489582,
+      "learning_rate": 0.0001,
+      "loss": 2.9229,
+      "loss/crossentropy": 2.3233034729957582,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2058998629450798,
+      "loss/reg": 0.0,
+      "step": 39240
+    },
+    {
+      "epoch": 0.2582236842105263,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.23391087849934897,
+      "learning_rate": 0.0001,
+      "loss": 2.9943,
+      "loss/crossentropy": 2.3962061643600463,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2253645345568657,
+      "loss/reg": 0.0,
+      "step": 39250
+    },
+    {
+      "epoch": 0.2582894736842105,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.0898577372233073,
+      "learning_rate": 0.0001,
+      "loss": 2.9846,
+      "loss/crossentropy": 2.2847093820571898,
+      "loss/hidden": 2.79375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20621878653764725,
+      "loss/reg": 0.0,
+      "step": 39260
+    },
+    {
+      "epoch": 0.25835526315789475,
+      "grad_norm": 2.859375,
+      "grad_norm_var": 0.062235514322916664,
+      "learning_rate": 0.0001,
+      "loss": 3.0113,
+      "loss/crossentropy": 2.1093714237213135,
+      "loss/hidden": 2.7484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22439000010490417,
+      "loss/reg": 0.0,
+      "step": 39270
+    },
+    {
+      "epoch": 0.25842105263157894,
+      "grad_norm": 2.625,
+      "grad_norm_var": 0.061091105143229164,
+      "learning_rate": 0.0001,
+      "loss": 2.9406,
+      "loss/crossentropy": 2.3419513583183287,
+      "loss/hidden": 2.625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2029005065560341,
+      "loss/reg": 0.0,
+      "step": 39280
+    },
+    {
+      "epoch": 0.2584868421052632,
+      "grad_norm": 1.96875,
+      "grad_norm_var": 0.0909332275390625,
+      "learning_rate": 0.0001,
+      "loss": 2.8985,
+      "loss/crossentropy": 2.230232834815979,
+      "loss/hidden": 2.690625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20411441028118132,
+      "loss/reg": 0.0,
+      "step": 39290
+    },
+    {
+      "epoch": 0.25855263157894737,
+      "grad_norm": 3.0625,
+      "grad_norm_var": 0.1186676025390625,
+      "learning_rate": 0.0001,
+      "loss": 2.9583,
+      "loss/crossentropy": 2.2822274684906008,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21612002551555634,
+      "loss/reg": 0.0,
+      "step": 39300
+    },
+    {
+      "epoch": 0.25861842105263155,
+      "grad_norm": 3.640625,
+      "grad_norm_var": 0.1718414306640625,
+      "learning_rate": 0.0001,
+      "loss": 3.0234,
+      "loss/crossentropy": 2.483612024784088,
+      "loss/hidden": 2.6703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22477086931467055,
+      "loss/reg": 0.0,
+      "step": 39310
+    },
+    {
+      "epoch": 0.2586842105263158,
+      "grad_norm": 2.578125,
+      "grad_norm_var": 0.18178609212239583,
+      "learning_rate": 0.0001,
+      "loss": 2.9349,
+      "loss/crossentropy": 2.3175944685935974,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2550591230392456,
+      "loss/reg": 0.0,
+      "step": 39320
+    },
+    {
+      "epoch": 0.25875,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.08434956868489583,
+      "learning_rate": 0.0001,
+      "loss": 2.8974,
+      "loss/crossentropy": 2.333976149559021,
+      "loss/hidden": 2.6578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20737813785672188,
+      "loss/reg": 0.0,
+      "step": 39330
+    },
+    {
+      "epoch": 0.2588157894736842,
+      "grad_norm": 2.65625,
+      "grad_norm_var": 0.0402984619140625,
+      "learning_rate": 0.0001,
+      "loss": 2.8992,
+      "loss/crossentropy": 1.9081893861293793,
+      "loss/hidden": 2.7,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19327801540493966,
+      "loss/reg": 0.0,
+      "step": 39340
+    },
+    {
+      "epoch": 0.2588815789473684,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.03901341756184896,
+      "learning_rate": 0.0001,
+      "loss": 2.8945,
+      "loss/crossentropy": 2.112404853105545,
+      "loss/hidden": 2.6296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1989109069108963,
+      "loss/reg": 0.0,
+      "step": 39350
+    },
+    {
+      "epoch": 0.25894736842105265,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.02379124959309896,
+      "learning_rate": 0.0001,
+      "loss": 2.8712,
+      "loss/crossentropy": 2.0947551012039183,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20299985110759736,
+      "loss/reg": 0.0,
+      "step": 39360
+    },
+    {
+      "epoch": 0.25901315789473683,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.08593343098958334,
+      "learning_rate": 0.0001,
+      "loss": 2.9641,
+      "loss/crossentropy": 2.160096913576126,
+      "loss/hidden": 2.7328125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23282611221075059,
+      "loss/reg": 0.0,
+      "step": 39370
+    },
+    {
+      "epoch": 0.2590789473684211,
+      "grad_norm": 2.46875,
+      "grad_norm_var": 0.09907124837239584,
+      "learning_rate": 0.0001,
+      "loss": 3.0059,
+      "loss/crossentropy": 2.1344396591186525,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20642745792865752,
+      "loss/reg": 0.0,
+      "step": 39380
+    },
+    {
+      "epoch": 0.25914473684210526,
+      "grad_norm": 2.546875,
+      "grad_norm_var": 0.0337066650390625,
+      "learning_rate": 0.0001,
+      "loss": 2.9501,
+      "loss/crossentropy": 2.211651337146759,
+      "loss/hidden": 2.575,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20566888749599457,
+      "loss/reg": 0.0,
+      "step": 39390
+    },
+    {
+      "epoch": 0.25921052631578945,
+      "grad_norm": 2.828125,
+      "grad_norm_var": 0.3254191080729167,
+      "learning_rate": 0.0001,
+      "loss": 2.9949,
+      "loss/crossentropy": 2.0947441935539244,
+      "loss/hidden": 2.771875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22096366733312606,
+      "loss/reg": 0.0,
+      "step": 39400
+    },
+    {
+      "epoch": 0.2592763157894737,
+      "grad_norm": 1.9921875,
+      "grad_norm_var": 0.19651667277018228,
+      "learning_rate": 0.0001,
+      "loss": 2.9392,
+      "loss/crossentropy": 2.3617849826812742,
+      "loss/hidden": 2.80625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25957188159227373,
+      "loss/reg": 0.0,
+      "step": 39410
+    },
+    {
+      "epoch": 0.2593421052631579,
+      "grad_norm": 2.328125,
+      "grad_norm_var": 0.0928179423014323,
+      "learning_rate": 0.0001,
+      "loss": 2.9103,
+      "loss/crossentropy": 2.373174512386322,
+      "loss/hidden": 2.7265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2120439499616623,
+      "loss/reg": 0.0,
+      "step": 39420
+    },
+    {
+      "epoch": 0.2594078947368421,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.053587849934895834,
+      "learning_rate": 0.0001,
+      "loss": 2.9039,
+      "loss/crossentropy": 2.2536337614059447,
+      "loss/hidden": 2.6078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18805657625198363,
+      "loss/reg": 0.0,
+      "step": 39430
+    },
+    {
+      "epoch": 0.2594736842105263,
+      "grad_norm": 2.0,
+      "grad_norm_var": 0.3204264322916667,
+      "learning_rate": 0.0001,
+      "loss": 3.0201,
+      "loss/crossentropy": 2.052464544773102,
+      "loss/hidden": 2.7265625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2089213587343693,
+      "loss/reg": 0.0,
+      "step": 39440
+    },
+    {
+      "epoch": 0.25953947368421054,
+      "grad_norm": 3.078125,
+      "grad_norm_var": 5.6367123205395795e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.1226,
+      "loss/crossentropy": 2.0904810786247254,
+      "loss/hidden": 2.53125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18436047285795212,
+      "loss/reg": 0.0,
+      "step": 39450
+    },
+    {
+      "epoch": 0.25960526315789473,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 5.636712321591454e+17,
+      "learning_rate": 0.0001,
+      "loss": 2.8972,
+      "loss/crossentropy": 2.383215081691742,
+      "loss/hidden": 2.7203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2062357634305954,
+      "loss/reg": 0.0,
+      "step": 39460
+    },
+    {
+      "epoch": 0.25967105263157897,
+      "grad_norm": 2.703125,
+      "grad_norm_var": 0.2125139872233073,
+      "learning_rate": 0.0001,
+      "loss": 3.0526,
+      "loss/crossentropy": 2.035197800397873,
+      "loss/hidden": 2.78125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22759751938283443,
+      "loss/reg": 0.0,
+      "step": 39470
+    },
+    {
+      "epoch": 0.25973684210526315,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.1047503153483073,
+      "learning_rate": 0.0001,
+      "loss": 2.8596,
+      "loss/crossentropy": 2.0758556723594666,
+      "loss/hidden": 2.5875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18243583887815476,
+      "loss/reg": 0.0,
+      "step": 39480
+    },
+    {
+      "epoch": 0.2598026315789474,
+      "grad_norm": 2.046875,
+      "grad_norm_var": 0.6220232645670573,
+      "learning_rate": 0.0001,
+      "loss": 2.9124,
+      "loss/crossentropy": 2.337350380420685,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2153509557247162,
+      "loss/reg": 0.0,
+      "step": 39490
+    },
+    {
+      "epoch": 0.2598684210526316,
+      "grad_norm": 2.28125,
+      "grad_norm_var": 0.24768651326497396,
+      "learning_rate": 0.0001,
+      "loss": 2.899,
+      "loss/crossentropy": 2.2104454159736635,
+      "loss/hidden": 2.671875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23622069656848907,
+      "loss/reg": 0.0,
+      "step": 39500
+    },
+    {
+      "epoch": 0.25993421052631577,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.06471354166666667,
+      "learning_rate": 0.0001,
+      "loss": 2.8843,
+      "loss/crossentropy": 2.006416219472885,
+      "loss/hidden": 2.7109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21008393466472625,
+      "loss/reg": 0.0,
+      "step": 39510
+    },
+    {
+      "epoch": 0.26,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 0.5674112955729167,
+      "learning_rate": 0.0001,
+      "loss": 3.0434,
+      "loss/crossentropy": 2.2104514479637145,
+      "loss/hidden": 2.7734375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.24286225736141204,
+      "loss/reg": 0.0,
+      "step": 39520
+    },
+    {
+      "epoch": 0.2600657894736842,
+      "grad_norm": 3.125,
+      "grad_norm_var": 2.339940388997396,
+      "learning_rate": 0.0001,
+      "loss": 2.9646,
+      "loss/crossentropy": 2.1471508383750915,
+      "loss/hidden": 2.875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2512055471539497,
+      "loss/reg": 0.0,
+      "step": 39530
+    },
+    {
+      "epoch": 0.26013157894736844,
+      "grad_norm": 2.53125,
+      "grad_norm_var": 1.9162261962890625,
+      "learning_rate": 0.0001,
+      "loss": 3.015,
+      "loss/crossentropy": 2.252353233098984,
+      "loss/hidden": 2.821875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22746440172195434,
+      "loss/reg": 0.0,
+      "step": 39540
+    },
+    {
+      "epoch": 0.2601973684210526,
+      "grad_norm": 2801795072.0,
+      "grad_norm_var": 4.9062847572554765e+17,
+      "learning_rate": 0.0001,
+      "loss": 3.0698,
+      "loss/crossentropy": 2.506903576850891,
+      "loss/hidden": 2.6703125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22553565353155136,
+      "loss/reg": 0.0,
+      "step": 39550
+    },
+    {
+      "epoch": 0.26026315789473686,
+      "grad_norm": 2.15625,
+      "grad_norm_var": 4.9062847573576256e+17,
+      "learning_rate": 0.0001,
+      "loss": 2.9221,
+      "loss/crossentropy": 2.4941019773483277,
+      "loss/hidden": 2.7796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23029526472091674,
+      "loss/reg": 0.0,
+      "step": 39560
+    },
+    {
+      "epoch": 0.26032894736842105,
+      "grad_norm": 2.609375,
+      "grad_norm_var": 0.12497456868489583,
+      "learning_rate": 0.0001,
+      "loss": 2.8649,
+      "loss/crossentropy": 2.1986557483673095,
+      "loss/hidden": 2.6109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.18305001109838487,
+      "loss/reg": 0.0,
+      "step": 39570
+    },
+    {
+      "epoch": 0.2603947368421053,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.10641276041666667,
+      "learning_rate": 0.0001,
+      "loss": 2.917,
+      "loss/crossentropy": 2.1990869998931886,
+      "loss/hidden": 2.6140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.1851473018527031,
+      "loss/reg": 0.0,
+      "step": 39580
+    },
+    {
+      "epoch": 0.2604605263157895,
+      "grad_norm": 2.375,
+      "grad_norm_var": 1.3978749593098958,
+      "learning_rate": 0.0001,
+      "loss": 3.0133,
+      "loss/crossentropy": 2.436345875263214,
+      "loss/hidden": 2.9578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.30376828759908675,
+      "loss/reg": 0.0,
+      "step": 39590
+    },
+    {
+      "epoch": 0.26052631578947366,
+      "grad_norm": 2.734375,
+      "grad_norm_var": 0.1521148681640625,
+      "learning_rate": 0.0001,
+      "loss": 2.9345,
+      "loss/crossentropy": 2.342786800861359,
+      "loss/hidden": 2.5828125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19189061373472213,
+      "loss/reg": 0.0,
+      "step": 39600
+    },
+    {
+      "epoch": 0.2605921052631579,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.15281575520833332,
+      "learning_rate": 0.0001,
+      "loss": 2.9767,
+      "loss/crossentropy": 2.4071336150169373,
+      "loss/hidden": 2.721875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22345878183841705,
+      "loss/reg": 0.0,
+      "step": 39610
+    },
+    {
+      "epoch": 0.2606578947368421,
+      "grad_norm": 1.9140625,
+      "grad_norm_var": 0.280962880452474,
+      "learning_rate": 0.0001,
+      "loss": 2.9462,
+      "loss/crossentropy": 2.1481564074754713,
+      "loss/hidden": 2.6421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20991018489003183,
+      "loss/reg": 0.0,
+      "step": 39620
+    },
+    {
+      "epoch": 0.26072368421052633,
+      "grad_norm": 3.609375,
+      "grad_norm_var": 0.6643674214680989,
+      "learning_rate": 0.0001,
+      "loss": 3.0336,
+      "loss/crossentropy": 2.0887478232383727,
+      "loss/hidden": 2.753125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22009581476449966,
+      "loss/reg": 0.0,
+      "step": 39630
+    },
+    {
+      "epoch": 0.2607894736842105,
+      "grad_norm": 2.21875,
+      "grad_norm_var": 0.36109619140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0258,
+      "loss/crossentropy": 2.138615000247955,
+      "loss/hidden": 2.765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2106044813990593,
+      "loss/reg": 0.0,
+      "step": 39640
+    },
+    {
+      "epoch": 0.26085526315789476,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.3787750244140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0633,
+      "loss/crossentropy": 1.9567537367343903,
+      "loss/hidden": 2.9484375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2987550154328346,
+      "loss/reg": 0.0,
+      "step": 39650
+    },
+    {
+      "epoch": 0.26092105263157894,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.15900065104166666,
+      "learning_rate": 0.0001,
+      "loss": 2.9489,
+      "loss/crossentropy": 2.524258053302765,
+      "loss/hidden": 2.7078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20932213813066483,
+      "loss/reg": 0.0,
+      "step": 39660
+    },
+    {
+      "epoch": 0.2609868421052632,
+      "grad_norm": 3.140625,
+      "grad_norm_var": 0.073583984375,
+      "learning_rate": 0.0001,
+      "loss": 2.9985,
+      "loss/crossentropy": 2.552958643436432,
+      "loss/hidden": 2.6203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21318122893571853,
+      "loss/reg": 0.0,
+      "step": 39670
+    },
+    {
+      "epoch": 0.26105263157894737,
+      "grad_norm": 2.359375,
+      "grad_norm_var": 0.0707672119140625,
+      "learning_rate": 0.0001,
+      "loss": 2.9692,
+      "loss/crossentropy": 2.2176095962524416,
+      "loss/hidden": 2.9203125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2208583801984787,
+      "loss/reg": 0.0,
+      "step": 39680
+    },
+    {
+      "epoch": 0.26111842105263156,
+      "grad_norm": 1.875,
+      "grad_norm_var": 0.058869425455729166,
+      "learning_rate": 0.0001,
+      "loss": 2.9243,
+      "loss/crossentropy": 2.2860087156295776,
+      "loss/hidden": 2.690625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2486639067530632,
+      "loss/reg": 0.0,
+      "step": 39690
+    },
+    {
+      "epoch": 0.2611842105263158,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.06942952473958333,
+      "learning_rate": 0.0001,
+      "loss": 2.9554,
+      "loss/crossentropy": 2.3581568241119384,
+      "loss/hidden": 2.834375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25539501309394835,
+      "loss/reg": 0.0,
+      "step": 39700
+    },
+    {
+      "epoch": 0.26125,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 1.3456217447916667,
+      "learning_rate": 0.0001,
+      "loss": 2.9286,
+      "loss/crossentropy": 2.494968020915985,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23517013043165208,
+      "loss/reg": 0.0,
+      "step": 39710
+    },
+    {
+      "epoch": 0.2613157894736842,
+      "grad_norm": 2.171875,
+      "grad_norm_var": 1.3488566080729167,
+      "learning_rate": 0.0001,
+      "loss": 3.0312,
+      "loss/crossentropy": 2.37692905664444,
+      "loss/hidden": 2.9,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.29073845595121384,
+      "loss/reg": 0.0,
+      "step": 39720
+    },
+    {
+      "epoch": 0.2613815789473684,
+      "grad_norm": 2.5,
+      "grad_norm_var": 0.202294921875,
+      "learning_rate": 0.0001,
+      "loss": 3.0294,
+      "loss/crossentropy": 2.281411385536194,
+      "loss/hidden": 2.7953125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22122962027788162,
+      "loss/reg": 0.0,
+      "step": 39730
+    },
+    {
+      "epoch": 0.26144736842105265,
+      "grad_norm": 2.015625,
+      "grad_norm_var": 0.2295318603515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9489,
+      "loss/crossentropy": 2.2690483570098876,
+      "loss/hidden": 2.9078125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2626466929912567,
+      "loss/reg": 0.0,
+      "step": 39740
+    },
+    {
+      "epoch": 0.26151315789473684,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.16734619140625,
+      "learning_rate": 0.0001,
+      "loss": 3.0073,
+      "loss/crossentropy": 2.394347441196442,
+      "loss/hidden": 2.709375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2102888874709606,
+      "loss/reg": 0.0,
+      "step": 39750
+    },
+    {
+      "epoch": 0.2615789473684211,
+      "grad_norm": 2.125,
+      "grad_norm_var": 0.2183990478515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9586,
+      "loss/crossentropy": 2.1268094956874846,
+      "loss/hidden": 2.728125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20743194743990898,
+      "loss/reg": 0.0,
+      "step": 39760
+    },
+    {
+      "epoch": 0.26164473684210526,
+      "grad_norm": 2.640625,
+      "grad_norm_var": 0.1597564697265625,
+      "learning_rate": 0.0001,
+      "loss": 2.9826,
+      "loss/crossentropy": 2.3416704416275023,
+      "loss/hidden": 2.6140625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20019679591059686,
+      "loss/reg": 0.0,
+      "step": 39770
+    },
+    {
+      "epoch": 0.26171052631578945,
+      "grad_norm": 3.03125,
+      "grad_norm_var": 0.08323160807291667,
+      "learning_rate": 0.0001,
+      "loss": 2.9672,
+      "loss/crossentropy": 2.0245181202888487,
+      "loss/hidden": 2.78125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21602422520518302,
+      "loss/reg": 0.0,
+      "step": 39780
+    },
+    {
+      "epoch": 0.2617763157894737,
+      "grad_norm": 2.515625,
+      "grad_norm_var": 0.12693583170572917,
+      "learning_rate": 0.0001,
+      "loss": 3.0371,
+      "loss/crossentropy": 2.2631885528564455,
+      "loss/hidden": 2.8890625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2282739758491516,
+      "loss/reg": 0.0,
+      "step": 39790
+    },
+    {
+      "epoch": 0.2618421052631579,
+      "grad_norm": 2.234375,
+      "grad_norm_var": 0.0588287353515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9485,
+      "loss/crossentropy": 2.1909173846244814,
+      "loss/hidden": 2.8171875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22533988654613496,
+      "loss/reg": 0.0,
+      "step": 39800
+    },
+    {
+      "epoch": 0.2619078947368421,
+      "grad_norm": 1.9765625,
+      "grad_norm_var": 0.07842178344726562,
+      "learning_rate": 0.0001,
+      "loss": 2.9508,
+      "loss/crossentropy": 2.3490309596061705,
+      "loss/hidden": 2.8765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2864062897861004,
+      "loss/reg": 0.0,
+      "step": 39810
+    },
+    {
+      "epoch": 0.2619736842105263,
+      "grad_norm": 2.8125,
+      "grad_norm_var": 0.08014500935872396,
+      "learning_rate": 0.0001,
+      "loss": 2.9507,
+      "loss/crossentropy": 2.5319572567939757,
+      "loss/hidden": 2.6375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2143532671034336,
+      "loss/reg": 0.0,
+      "step": 39820
+    },
+    {
+      "epoch": 0.26203947368421054,
+      "grad_norm": 1.96875,
+      "grad_norm_var": 0.0728515625,
+      "learning_rate": 0.0001,
+      "loss": 2.9447,
+      "loss/crossentropy": 2.350716245174408,
+      "loss/hidden": 2.5578125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21392515301704407,
+      "loss/reg": 0.0,
+      "step": 39830
+    },
+    {
+      "epoch": 0.26210526315789473,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.07763264973958334,
+      "learning_rate": 0.0001,
+      "loss": 2.9706,
+      "loss/crossentropy": 2.2897894740104676,
+      "loss/hidden": 2.73125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.242967389523983,
+      "loss/reg": 0.0,
+      "step": 39840
+    },
+    {
+      "epoch": 0.26217105263157897,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.138525390625,
+      "learning_rate": 0.0001,
+      "loss": 3.0426,
+      "loss/crossentropy": 2.309167265892029,
+      "loss/hidden": 2.9640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2831707686185837,
+      "loss/reg": 0.0,
+      "step": 39850
+    },
+    {
+      "epoch": 0.26223684210526316,
+      "grad_norm": 2.453125,
+      "grad_norm_var": 0.13242085774739584,
+      "learning_rate": 0.0001,
+      "loss": 3.0234,
+      "loss/crossentropy": 2.3926889896392822,
+      "loss/hidden": 2.5875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.19550751000642777,
+      "loss/reg": 0.0,
+      "step": 39860
+    },
+    {
+      "epoch": 0.26230263157894734,
+      "grad_norm": 2.265625,
+      "grad_norm_var": 0.028873697916666666,
+      "learning_rate": 0.0001,
+      "loss": 2.8699,
+      "loss/crossentropy": 2.389265012741089,
+      "loss/hidden": 2.5765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20654226988554,
+      "loss/reg": 0.0,
+      "step": 39870
+    },
+    {
+      "epoch": 0.2623684210526316,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.04534098307291667,
+      "learning_rate": 0.0001,
+      "loss": 2.9553,
+      "loss/crossentropy": 2.052034729719162,
+      "loss/hidden": 2.7765625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20443758517503738,
+      "loss/reg": 0.0,
+      "step": 39880
+    },
+    {
+      "epoch": 0.26243421052631577,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.06367899576822916,
+      "learning_rate": 0.0001,
+      "loss": 2.8898,
+      "loss/crossentropy": 2.276283013820648,
+      "loss/hidden": 2.78125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.25378709435462954,
+      "loss/reg": 0.0,
+      "step": 39890
+    },
+    {
+      "epoch": 0.2625,
+      "grad_norm": 1.9609375,
+      "grad_norm_var": 0.06096369425455729,
+      "learning_rate": 0.0001,
+      "loss": 2.9238,
+      "loss/crossentropy": 2.3180172562599184,
+      "loss/hidden": 2.7109375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23101864904165267,
+      "loss/reg": 0.0,
+      "step": 39900
+    },
+    {
+      "epoch": 0.2625657894736842,
+      "grad_norm": 2.34375,
+      "grad_norm_var": 0.1415911356608073,
+      "learning_rate": 0.0001,
+      "loss": 2.9487,
+      "loss/crossentropy": 2.6349021911621096,
+      "loss/hidden": 3.0421875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.27188998460769653,
+      "loss/reg": 0.0,
+      "step": 39910
+    },
+    {
+      "epoch": 0.26263157894736844,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.11770731608072917,
+      "learning_rate": 0.0001,
+      "loss": 2.8688,
+      "loss/crossentropy": 2.276544678211212,
+      "loss/hidden": 2.7640625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20636183619499207,
+      "loss/reg": 0.0,
+      "step": 39920
+    },
+    {
+      "epoch": 0.2626973684210526,
+      "grad_norm": 2.1875,
+      "grad_norm_var": 0.1878814697265625,
+      "learning_rate": 0.0001,
+      "loss": 3.023,
+      "loss/crossentropy": 2.388458263874054,
+      "loss/hidden": 2.7296875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23089798912405968,
+      "loss/reg": 0.0,
+      "step": 39930
+    },
+    {
+      "epoch": 0.26276315789473687,
+      "grad_norm": 2.484375,
+      "grad_norm_var": 0.14111226399739582,
+      "learning_rate": 0.0001,
+      "loss": 2.9919,
+      "loss/crossentropy": 2.0341047286987304,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.21918695122003556,
+      "loss/reg": 0.0,
+      "step": 39940
+    },
+    {
+      "epoch": 0.26282894736842105,
+      "grad_norm": 3.1875,
+      "grad_norm_var": 0.183544921875,
+      "learning_rate": 0.0001,
+      "loss": 3.0396,
+      "loss/crossentropy": 2.3217403292655945,
+      "loss/hidden": 2.95625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.26592089235782623,
+      "loss/reg": 0.0,
+      "step": 39950
+    },
+    {
+      "epoch": 0.26289473684210524,
+      "grad_norm": 2.140625,
+      "grad_norm_var": 0.0977691650390625,
+      "learning_rate": 0.0001,
+      "loss": 2.9174,
+      "loss/crossentropy": 2.55605833530426,
+      "loss/hidden": 2.690625,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2847170978784561,
+      "loss/reg": 0.0,
+      "step": 39960
+    },
+    {
+      "epoch": 0.2629605263157895,
+      "grad_norm": 2.5625,
+      "grad_norm_var": 0.06572977701822917,
+      "learning_rate": 0.0001,
+      "loss": 2.9685,
+      "loss/crossentropy": 1.9914783239364624,
+      "loss/hidden": 2.784375,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.20305539071559905,
+      "loss/reg": 0.0,
+      "step": 39970
+    },
+    {
+      "epoch": 0.26302631578947366,
+      "grad_norm": 2.25,
+      "grad_norm_var": 0.04635416666666667,
+      "learning_rate": 0.0001,
+      "loss": 2.9532,
+      "loss/crossentropy": 2.478356397151947,
+      "loss/hidden": 2.8796875,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.2187245175242424,
+      "loss/reg": 0.0,
+      "step": 39980
+    },
+    {
+      "epoch": 0.2630921052631579,
+      "grad_norm": 2.3125,
+      "grad_norm_var": 0.03802083333333333,
+      "learning_rate": 0.0001,
+      "loss": 2.9101,
+      "loss/crossentropy": 2.4963066220283507,
+      "loss/hidden": 2.678125,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.23317703753709793,
+      "loss/reg": 0.0,
+      "step": 39990
+    },
+    {
+      "epoch": 0.2631578947368421,
+      "grad_norm": 2.6875,
+      "grad_norm_var": 0.10871480305989584,
+      "learning_rate": 0.0001,
+      "loss": 2.9937,
+      "loss/crossentropy": 2.340533971786499,
+      "loss/hidden": 2.7,
+      "loss/incoh": 0.0,
+      "loss/logits": 0.22518680542707442,
+      "loss/reg": 0.0,
+      "step": 40000
+    }
+  ],
+  "logging_steps": 10,
+  "max_steps": 152000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 9223372036854775807,
+  "save_steps": 20000,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 2.8575100320088064e+19,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}