library_name, tags, datasets, model-index

library_name	tags	datasets	model-index
transformers	alignment-handbook epsilon-dpo generated_from_trainer	Anthropic/hh-rlhf	name results qwen3-8b-base-epsilon-dpo-hh-harmless-4xh200-batch-64-20260418-172841

qwen3-8b-base-epsilon-dpo-hh-harmless-4xh200-batch-64-20260418-172841

This model is a fine-tuned version of /scratch/qu.yang1/dynamic-dpo-v4/outputs/qwen3-8b-base-sft-hh-helpful-4xh200-batch-64-20260417-214452 on the Anthropic/hh-rlhf dataset. It achieves the following results on the evaluation set:

• Loss: 0.5753
• Epsilon Dpo/beta: 0.0120
• Epsilon Dpo/loss Margin Mean: 30.1703
• Epsilon Dpo/beta Margin Mean: 0.3587
• Epsilon Dpo/beta Margin Std: 0.6142
• Epsilon Dpo/beta Margin Grad Mean: -0.4187
• Epsilon Dpo/beta Margin Grad Std: 0.1376
• Rewards/chosen: -0.5348
• Rewards/rejected: -0.8935
• Rewards/accuracies: 0.7236
• Rewards/margins: 0.3587
• Logps/chosen: -131.7627
• Logps/rejected: -178.7413
• Logps/ref Chosen: -87.4272
• Logps/ref Rejected: -104.2355
• Logits/chosen: -0.2089
• Logits/rejected: -0.3424
• Kl/p Epsilon Steps: 0.7170
• Kl/n Epsilon Steps: 0.2817

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
1.3499	0.1512	100	0.6692	0.0877	0.6693	0.0574	0.1823	-0.4858	0.0449	-0.0275	-0.0849	0.6312	0.0574	-87.7332	-105.2108	-87.4272	-104.2355	0.1122	0.1240	0.6241	0.3759
1.1283	0.3023	200	0.6028	0.0636	4.8746	0.3057	0.6498	-0.4313	0.1445	-0.1604	-0.4661	0.6831	0.3057	-89.9090	-111.5920	-87.4272	-104.2355	-0.3390	-0.3863	0.6734	0.3261
1.2376	0.4535	300	0.5699	0.0434	10.8082	0.4637	0.8475	-0.4021	0.1759	-0.4198	-0.8835	0.7086	0.4637	-97.0338	-124.6503	-87.4272	-104.2355	-0.3510	-0.4271	0.6941	0.3050
1.2438	0.6047	400	0.5624	0.0290	16.4413	0.4712	0.8253	-0.4003	0.1706	-0.4065	-0.8777	0.7179	0.4712	-101.3740	-134.6236	-87.4272	-104.2355	-0.3535	-0.4572	0.7064	0.2923
1.1147	0.7559	500	0.5588	0.0187	26.6232	0.4919	0.8457	-0.3960	0.1755	-0.8059	-1.2978	0.7205	0.4919	-130.4120	-173.8436	-87.4272	-104.2355	-0.1937	-0.3318	0.7157	0.2843
1.1925	0.9070	600	0.5753	0.0120	30.1703	0.3587	0.6142	-0.4187	0.1376	-0.5348	-0.8935	0.7236	0.3587	-131.7627	-178.7413	-87.4272	-104.2355	-0.2089	-0.3424	0.7170	0.2817

Framework versions

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