llama-3-8b-base-new-dpo-hh-harmless-s_star1.0-4xh200-batch-64-20260421-213851

• Loss: 0.5467
• Fcm Dpo/beta: 0.2268
• Fcm Dpo/q T: 0.3412
• Fcm Dpo/delta: -0.0017
• Fcm Dpo/margin: 4.4089
• Margin Dpo/margin Mean: 4.4089
• Margin Dpo/margin Std: 7.2504
• Logps/chosen: -82.5570
• Logps/rejected: -91.6554
• Logps/ref Chosen: -74.8595
• Logps/ref Rejected: -79.5490
• Logits/chosen: 0.2724
• Logits/rejected: 0.2290

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

• 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