mistral-7b-base-epsilon-dpo-hh-harmless-4xh200-batch-64/README.md

library_name, base_model, tags, datasets, model-index

library_name	base_model	tags	datasets	model-index
transformers	mistral-7b-base-sft-hh-harmless-4xh200-batch-64-20260418-015332	alignment-handbook epsilon-dpo generated_from_trainer	Anthropic/hh-rlhf	name results mistral-7b-base-epsilon-dpo-hh-harmless-4xh200-batch-64-20260418-015332

mistral-7b-base-epsilon-dpo-hh-harmless-4xh200-batch-64-20260418-015332

This model is a fine-tuned version of mistral-7b-base-sft-hh-harmless-4xh200-batch-64-20260418-015332 on the Anthropic/hh-rlhf dataset. It achieves the following results on the evaluation set:

• Loss: 0.5935
• Epsilon Dpo/beta: 0.0044
• Epsilon Dpo/loss Margin Mean: 65.2805
• Epsilon Dpo/beta Margin Mean: 0.2810
• Epsilon Dpo/beta Margin Std: 0.5001
• Epsilon Dpo/beta Margin Grad Mean: -0.4338
• Epsilon Dpo/beta Margin Grad Std: 0.1154
• Rewards/chosen: -0.4127
• Rewards/rejected: -0.6936
• Rewards/accuracies: 0.7196
• Rewards/margins: 0.2810
• Logps/chosen: -171.7436
• Logps/rejected: -233.1436
• Logps/ref Chosen: -77.4087
• Logps/ref Rejected: -73.5282
• Logits/chosen: -3.1674
• Logits/rejected: -2.9705
• Kl/p Epsilon Steps: 0.7117
• Kl/n Epsilon Steps: 0.2870

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-07
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• total_eval_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Epsilon Dpo/beta	Epsilon Dpo/loss Margin Mean	Epsilon Dpo/beta Margin Mean	Epsilon Dpo/beta Margin Std	Epsilon Dpo/beta Margin Grad Mean	Epsilon Dpo/beta Margin Grad Std	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/chosen	Logps/rejected	Logps/ref Chosen	Logps/ref Rejected	Logits/chosen	Logits/rejected	Kl/p Epsilon Steps	Kl/n Epsilon Steps
0.9822	0.1512	100	0.6127	0.0682	6.4966	0.4365	1.0172	-0.4126	0.1971	-0.5724	-1.0089	0.6646	0.4365	-85.7551	-88.3712	-77.4087	-73.5282	-3.5599	-3.5776	0.6721	0.3275
0.8262	0.3023	200	0.5655	0.0393	16.0515	0.6241	1.1422	-0.3816	0.2119	-0.8102	-1.4343	0.7007	0.6241	-97.9143	-110.0853	-77.4087	-73.5282	-3.5815	-3.5644	0.7073	0.2918
0.949	0.4535	300	0.5616	0.0224	23.5752	0.5225	0.9256	-0.3915	0.1875	-0.6680	-1.1905	0.7121	0.5225	-107.0630	-126.7577	-77.4087	-73.5282	-3.5920	-3.5350	0.7069	0.2918
0.9633	0.6047	400	0.5351	0.0130	47.5045	0.6097	0.9587	-0.3752	0.1913	-0.7854	-1.3951	0.7267	0.6097	-137.7568	-181.3808	-77.4087	-73.5282	-3.4438	-3.3090	0.7284	0.2698
1.0329	0.7559	500	0.5608	0.0073	59.4602	0.4316	0.7150	-0.4038	0.1553	-0.5800	-1.0116	0.7258	0.4316	-156.1025	-211.6822	-77.4087	-73.5282	-3.2563	-3.0824	0.7192	0.2790
1.0833	0.9070	600	0.5935	0.0044	65.2805	0.2810	0.5001	-0.4338	0.1154	-0.4127	-0.6936	0.7196	0.2810	-171.7436	-233.1436	-77.4087	-73.5282	-3.1674	-2.9705	0.7117	0.2870

Framework versions

• Transformers 4.51.0
• Pytorch 2.3.1+cu121
• Datasets 2.21.0
• Tokenizers 0.21.4