mistral-7b-base-epsilon-dpo-hh-helpful-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-helpful-4xh200-batch-64-20260418-015332	alignment-handbook epsilon-dpo generated_from_trainer	Anthropic/hh-rlhf	name results mistral-7b-base-epsilon-dpo-hh-helpful-4xh200-batch-64-20260418-015332

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

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

• Loss: 0.5823
• Epsilon Dpo/beta: 0.0009
• Epsilon Dpo/loss Margin Mean: 478.8302
• Epsilon Dpo/beta Margin Mean: 0.4461
• Epsilon Dpo/beta Margin Std: 0.8791
• Epsilon Dpo/beta Margin Grad Mean: -0.4072
• Epsilon Dpo/beta Margin Grad Std: 0.1767
• Rewards/chosen: -0.6675
• Rewards/rejected: -1.1135
• Rewards/accuracies: 0.7038
• Rewards/margins: 0.4461
• Logps/chosen: -771.9094
• Logps/rejected: -1274.5287
• Logps/ref Chosen: -68.2911
• Logps/ref Rejected: -92.0804
• Logits/chosen: -3.8786
• Logits/rejected: -4.7234
• Kl/p Epsilon Steps: 0.6819
• Kl/n Epsilon Steps: 0.3176

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
0.6663	0.1468	100	0.4936	0.0513	29.5905	1.5046	2.1537	-0.3036	0.2585	-1.4541	-2.9587	0.7556	1.5046	-96.5378	-149.9176	-68.2911	-92.0804	-3.2766	-3.2684	0.7226	0.2765
0.559	0.2937	200	0.4014	0.0224	78.6564	1.7522	1.9999	-0.2627	0.2420	-2.4182	-4.1704	0.7984	1.7522	-175.9333	-278.3791	-68.2911	-92.0804	-2.2068	-2.1055	0.7821	0.2175
0.4996	0.4405	300	0.4131	0.0097	122.2527	1.1844	1.3051	-0.2961	0.2077	-1.5383	-2.7227	0.7928	1.1844	-225.6753	-371.7173	-68.2911	-92.0804	-1.9427	-1.7790	0.7783	0.2217
0.5665	0.5874	400	0.4863	0.0042	180.0219	0.7587	0.9839	-0.3501	0.1877	-1.0773	-1.8360	0.7564	0.7587	-320.8791	-524.6903	-68.2911	-92.0804	-2.2465	-2.1282	0.7419	0.2581
0.8394	0.7342	500	0.5502	0.0019	226.7985	0.4281	0.6344	-0.4041	0.1391	-0.6045	-1.0326	0.7453	0.4281	-384.0945	-634.6824	-68.2911	-92.0804	-2.9733	-3.1795	0.7290	0.2701
0.8895	0.8811	600	0.5823	0.0009	478.8302	0.4461	0.8791	-0.4072	0.1767	-0.6675	-1.1135	0.7038	0.4461	-771.9094	-1274.5287	-68.2911	-92.0804	-3.8786	-4.7234	0.6819	0.3176

Framework versions

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