args.json

{
    "project": "iterative preference learning",
    "exp_name": "llama-3-8b-offline-dpo",
    "cache_dir": ".cache",
    "result_dir": "results",
    "data": null,
    "prompt_max_length": 1024,
    "max_length": 2048,
    "model_name_or_path": "RLHFlow/LLaMA3-SFT",
    "ref_model_name_or_path": "RLHFlow/LLaMA3-SFT",
    "beta": 0.1,
    "n_epochs": 1,
    "per_device_batch_size": 1,
    "gradient_accumulation_steps": 32,
    "lr": 5e-07,
    "warmup_ratio": 0.03,
    "max_grad_norm": 1,
    "open_port": 60843
}