mistral-7b-base-margin-dpo-hh-helpful-4xh200-batch-64

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

mistral-7b-base-margin-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: 0.3349
• Margin Dpo/beta: 0.1000
• Margin Dpo/loss Margin Mean: 41.5652
• Margin Dpo/beta Margin Mean: 4.1565
• Margin Dpo/beta Margin Std: 4.1510
• Margin Dpo/beta Margin Grad Mean: -0.1899
• Margin Dpo/beta Margin Grad Std: 0.2630
• Margin Dpo/margin Mean: 41.5652
• Margin Dpo/margin Std: 41.5097
• Logps/chosen: -113.7768
• Logps/rejected: -179.1313
• Logps/ref Chosen: -68.2911
• Logps/ref Rejected: -92.0804
• Logits/chosen: -2.5186
• Logits/rejected: -2.4962

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

Framework versions

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