qwen3-8b-base-new-dpo-ultrafeedback-4xh200-batch-128-q_t-0.4-s_star-0.35-20260430-140517

• Loss: 0.6076
• Fcm Dpo/beta: 0.0040
• Margin Dpo/margin Mean: 54.4214
• Margin Dpo/margin Std: 86.7093
• Logps/chosen: -331.5330
• Logps/rejected: -370.1833
• Logps/ref Chosen: -280.4167
• Logps/ref Rejected: -264.6455
• Kl/chosen Kl Mean: -51.1163
• Kl/rejected Kl Mean: -105.5378
• Kl/mean: -78.3270
• Kl/std: 74.4785
• Logits/chosen: 1.4424
• Logits/rejected: 1.5060

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

• learning_rate: 5e-07
• train_batch_size: 4
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• total_eval_batch_size: 8
• 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