{
    "actor_model_name_or_path": "output/sft_opt_1b",
    "reward_model_name_or_path": "output/rm_opt",
    "cost_model_name_or_path": "output/cm_opt",
    "reward_critic_model_name_or_path": "output/rm_opt",
    "cost_critic_model_name_or_path": "output/cm_opt",
    "max_length": 512,
    "trust_remote_code": true,
    "train_datasets": [
        [
            "PKU-SafeRLHF/train",
            {
                "proportion": 1.0
            }
        ]
    ],
    "ptx_datasets": [
        [
            "alpaca",
            {
                "proportion": 1.0
            }
        ]
    ],
    "eval_datasets": null,
    "threshold": 0.0,
    "lambda_init": 1.0,
    "lambda_lr": 0.1,
    "lambda_max": 5.0,
    "lambda_update_delay_steps": 0,
    "episode_cost_window_size": 128,
    "kl_coeff": 0.01,
    "clip_range_ratio": 0.2,
    "clip_range_score": 50.0,
    "clip_range_value": 5.0,
    "ptx_coeff": 16.0,
    "epochs": 1,
    "update_iters": 1,
    "per_device_prompt_batch_size": 16,
    "per_device_train_batch_size": 16,
    "per_device_eval_batch_size": 16,
    "gradient_accumulation_steps": 2,
    "actor_lr": 1e-05,
    "actor_weight_decay": 0.01,
    "actor_lr_scheduler_type": "cosine",
    "actor_lr_warmup_ratio": 0.03,
    "actor_gradient_checkpointing": true,
    "critic_lr": 5e-06,
    "critic_weight_decay": 0.0,
    "critic_lr_scheduler_type": "constant",
    "critic_lr_warmup_ratio": 0.03,
    "critic_gradient_checkpointing": true,
    "normalize_reward": false,
    "normalize_cost": false,
    "seed": 42,
    "fp16": true,
    "bf16": false,
    "tf32": null,
    "temperature": 1.0,
    "top_p": 1.0,
    "num_return_sequences": 1,
    "repetition_penalty": 1.0,
    "eval_strategy": "epoch",
    "eval_interval": 1000000,
    "need_eval": false,
    "eval_split_ratio": null,
    "output_dir": "/home/rwliang/Reinforcement-Learning/FINAL_PROJECT/rl-final/safe-rlhf/output/ppo-lag",
    "log_type": "wandb",
    "log_dir": "/home/rwliang/Reinforcement-Learning/FINAL_PROJECT/rl-final/safe-rlhf/output/ppo-lag",
    "log_project": "Safe-RLHF-PPO",
    "log_run_name": "ppo_lag-2024-12-04-22-29-24",
    "save_16bit": false,
    "save_interval": 1000000,
    "local_rank": 0,
    "zero_stage": 1,
    "offload": "none",
    "deepspeed": false,
    "deepspeed_config": null,
    "deepscale": false,
    "deepscale_config": null,
    "global_rank": 0,
    "device": {
        "type": "torch.device",
        "repr": "device(type='cuda', index=0)"
    },
    "total_training_steps": 806
}