jackf857/qwen3-8b-base-beta-dpo-hh-harmless-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. -->

qwen3-8b-base-beta-dpo-hh-harmless-4xh200-batch-64-20260418-012645

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: 0.6518
• —Beta Dpo/beta: 0.1628
• —Beta Dpo/loss Margin Mean: 8.6336
• —Beta Dpo/beta Margin Mean: 1.6614
• —Beta Dpo/beta Margin Std: 2.5181
• —Beta Dpo/beta Margin Grad Mean: -0.3428
• —Beta Dpo/beta Margin Grad Std: 0.2482
• —Beta Dpo/gap Mean: 7.7476
• —Beta Dpo/gap Std: 14.8995
• —Beta Dpo/beta Used Raw: 0.1532
• —Beta Dpo/beta Used: 0.1628
• —Beta Dpo/mask Keep Frac: 1.0
• —Logits/chosen: 1.2071
• —Logits/rejected: 1.1403

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
• —trainbatchsize: 8
• —evalbatchsize: 8
• —seed: 42
• —distributed_type: multi-GPU
• —num_devices: 4
• —gradientaccumulationsteps: 2
• —totaltrainbatch_size: 64
• —totalevalbatch_size: 32
• —optimizer: Use OptimizerNames.ADAMWTORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizerargs=No additional optimizer arguments
• —lrschedulertype: cosine
• —lrschedulerwarmup_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