v1_1000_STEPS_1e7_rate_05_beta_DPO

This model is a fine-tuned version of mistralai/Mistral-7B-Instruct-v0.1 on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.6349
• Rewards/chosen: -0.2276
• Rewards/rejected: -0.4095
• Rewards/accuracies: 0.5890
• Rewards/margins: 0.1819
• Logps/rejected: -17.6986
• Logps/chosen: -15.7083
• Logits/rejected: -3.3433
• Logits/chosen: -3.3435

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: 1e-07
• train_batch_size: 2
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• training_steps: 1000

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6758	0.1	100	0.6807	-0.0108	-0.0388	0.5582	0.0280	-16.9571	-15.2746	-3.3527	-3.3528
0.648	0.2	200	0.6605	-0.0898	-0.1746	0.5692	0.0849	-17.2288	-15.4326	-3.3470	-3.3471
0.6324	0.29	300	0.6498	-0.1892	-0.3115	0.5802	0.1224	-17.5026	-15.6314	-3.3449	-3.3450
0.6949	0.39	400	0.6438	-0.1429	-0.2881	0.5912	0.1452	-17.4557	-15.5388	-3.3451	-3.3452
0.6848	0.49	500	0.6369	-0.1735	-0.3420	0.6066	0.1685	-17.5635	-15.6000	-3.3438	-3.3439
0.6344	0.59	600	0.6375	-0.2102	-0.3842	0.5846	0.1740	-17.6480	-15.6735	-3.3436	-3.3437
0.6551	0.68	700	0.6366	-0.2240	-0.4017	0.5868	0.1777	-17.6829	-15.7010	-3.3433	-3.3434
0.5891	0.78	800	0.6356	-0.2274	-0.4088	0.6066	0.1813	-17.6971	-15.7079	-3.3433	-3.3434
0.6461	0.88	900	0.6348	-0.2270	-0.4096	0.5956	0.1826	-17.6988	-15.7070	-3.3433	-3.3435
0.6059	0.98	1000	0.6349	-0.2276	-0.4095	0.5890	0.1819	-17.6986	-15.7083	-3.3433	-3.3435

Framework versions

• Transformers 4.39.1
• Pytorch 2.0.0+cu117
• Datasets 2.18.0
• Tokenizers 0.15.2