mistral-7b-expo-7b-L2EXPO-25-7

This model is a fine-tuned version of hZzy/mistral-7b-sft-25-1 on the hZzy/direction_right2 dataset. It achieves the following results on the evaluation set:

• Loss: 0.4452
• Objective: 0.4429
• Reward Accuracy: 0.6574
• Logp Accuracy: 0.6451
• Log Diff Policy: 17.7719
• Chosen Logps: -193.5036
• Rejected Logps: -211.2754
• Chosen Rewards: -0.9882
• Rejected Rewards: -1.1621
• Logits: -1.8933

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-06
• train_batch_size: 3
• eval_batch_size: 3
• seed: 42
• distributed_type: multi-GPU
• num_devices: 3
• gradient_accumulation_steps: 12
• total_train_batch_size: 108
• total_eval_batch_size: 9
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant_with_warmup
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Objective	Reward Accuracy	Logp Accuracy	Log Diff Policy	Chosen Logps	Rejected Logps	Chosen Rewards	Rejected Rewards	Logits
0.6123	0.1517	100	0.5111	0.5082	0.5386	0.5193	0.4932	-92.5590	-93.0521	0.0212	0.0201	-2.1823
0.586	0.3033	200	0.4950	0.4930	0.5886	0.5540	2.7175	-94.1884	-96.9058	0.0049	-0.0184	-2.0328
0.5397	0.4550	300	0.4754	0.4737	0.6337	0.6169	10.7924	-129.7178	-140.5102	-0.3504	-0.4545	-1.8133
0.492	0.6067	400	0.4631	0.4616	0.6485	0.6401	13.2720	-123.6336	-136.9056	-0.2895	-0.4184	-2.1057
0.5057	0.7583	500	0.4569	0.4552	0.6569	0.6530	16.0166	-157.9163	-173.9330	-0.6323	-0.7887	-2.0772
0.4562	0.9100	600	0.4556	0.4556	0.6683	0.6639	17.9609	-185.6944	-203.6553	-0.9101	-1.0859	-2.0650
0.4627	1.0617	700	0.4463	0.4468	0.6683	0.6580	17.2045	-193.8147	-211.0192	-0.9913	-1.1595	-2.0785
0.4675	1.2133	800	0.4477	0.4481	0.6667	0.6586	18.6212	-171.0579	-189.6790	-0.7638	-0.9461	-2.0805
0.4631	1.3650	900	0.4523	0.4507	0.6697	0.6614	19.3812	-184.6346	-204.0157	-0.8995	-1.0895	-2.1364
0.4706	1.5167	1000	0.4428	0.4430	0.6636	0.6535	18.2421	-175.5473	-193.7894	-0.8087	-0.9872	-2.0802
0.4404	1.6684	1100	0.4509	0.4525	0.6644	0.6544	19.2144	-205.9645	-225.1790	-1.1128	-1.3011	-2.0013
0.4086	1.8200	1200	0.4418	0.4425	0.6742	0.6611	20.1585	-203.0412	-223.1997	-1.0836	-1.2814	-2.0214
0.4211	1.9717	1300	0.4377	0.4377	0.6636	0.6488	17.4303	-204.1001	-221.5304	-1.0942	-1.2647	-1.9872
0.3854	2.1234	1400	0.4415	0.4413	0.6616	0.6521	17.5702	-221.2297	-238.7999	-1.2655	-1.4374	-1.9931
0.4044	2.2750	1500	0.4484	0.4486	0.6644	0.6527	20.0957	-200.4279	-220.5237	-1.0575	-1.2546	-1.9230
0.4357	2.4267	1600	0.4485	0.4484	0.6703	0.6600	20.6916	-175.1429	-195.8344	-0.8046	-1.0077	-1.9267
0.4092	2.5784	1700	0.4627	0.4633	0.6641	0.6555	22.5421	-193.3455	-215.8876	-0.9866	-1.2082	-1.8966
0.4004	2.7300	1800	0.4512	0.4522	0.6625	0.6516	19.5557	-202.4462	-222.0019	-1.0776	-1.2694	-1.8232
0.3783	2.8817	1900	0.4530	0.4514	0.6630	0.6485	20.0630	-199.5361	-219.5991	-1.0485	-1.2453	-1.8925

Framework versions

• PEFT 0.11.1
• Transformers 4.42.0
• Pytorch 2.6.0+cu124
• Datasets 3.2.0
• Tokenizers 0.19.1