Saving weights and logs of step 10000

events.out.tfevents.1642108860.t1v-n-6f5efcd5-w-0.604413.0.v2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ab5fc1b7b81c77898124077e503b27d796784c21a8862243e091545846c1404
+size 1470136

flax_model.msgpack

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ef44b87241742f4f804608e53a522ca494c43160f37a11baebfe8203d1b82c3a
+oid sha256:7cdf5da2a3aae0b2ad6669aebb07d2b02b88c0103039ec07d2220717a4667368
 size 498796983

run_mlm_flax.py

@@ -129,6 +129,10 @@ class DataTrainingArguments:
     static_learning_rate: bool = field(
         default=False, metadata={"help": "Use a non decaying learning rate"}
     )
+    end_learning_rate: float = field(
+        default=0, metadata={"help": "End learning rate. Will be ignored it startic learning rate is set"}
+    )
+    
     auth_token: bool = field(
         default=False, metadata={"help": "Use authorisation token"}
     )
@@ -510,6 +514,8 @@ if __name__ == "__main__":
 
     if data_args.static_learning_rate:
         end_lr_value = training_args.learning_rate
+    elif data_args.end_learning_rate:
+        end_lr_value = data_args.end_learning_rate
     else:
         end_lr_value = 0
 

run_step5.sh

@@ -0,0 +1,27 @@
+./run_mlm_flax.py \
+    --output_dir="./" \
+    --model_type="roberta" \
+    --model_name_or_path="./" \
+    --config_name="./" \
+    --tokenizer_name="./" \
+    --train_file /mnt/disks/flaxdisk/corpus/train_1_4.json \
+    --validation_file /mnt/disks/flaxdisk/corpus/validation.json \
+    --cache_dir="/mnt/disks/flaxdisk/cache/" \
+    --max_seq_length="128" \
+    --weight_decay="0.01" \
+    --per_device_train_batch_size="200" \
+    --per_device_eval_batch_size="200" \
+    --learning_rate="6e-4" \
+    --warmup_steps="0" \
+    --overwrite_output_dir \
+    --num_train_epochs="6" \
+    --adam_beta1="0.9" \
+    --adam_beta2="0.98" \
+    --logging_steps="10000" \
+    --save_steps="10000" \
+    --eval_steps="10000" \
+    --preprocessing_num_workers="64" \
+    --auth_token="True" \
+    --dtype="bfloat16" \
+    --adafactor \
+    --push_to_hub