{
    "cmd": "dpo",
    "requirements":{
        "gpu": "8",
        "ddp": "8"
    },
    "eval_requirements": {
      "gpu": "1"
    },
    "eval_dataset": ["ceval", "gsm8k", "arc"],
    "args": {
        "model_type": "llama2-7b",
        "ref_model_type": "llama2-7b",
        "template_type": "llama",
        "dataset": "hh-rlhf-cn",
        "train_dataset_sample": 200000,
        "truncation_strategy": "truncation_left",
        "val_dataset_sample": 10000,
        "num_train_epochs": 1,
        "max_length": 1024,
        "max_prompt_length": 512,
        "check_dataset_strategy": "none",
        "gradient_checkpointing": true,
        "batch_size": 1,
        "weight_decay": 0.01,
        "learning_rate": 5e-5,
        "gradient_accumulation_steps": 2,
        "max_grad_norm": 1.0,
        "warmup_ratio": 0.03,
        "eval_steps": 2000,
        "save_steps": 2000,
        "save_total_limit": 2,
        "logging_steps": 10,
        "sft_type": "lora",
        "lora_target_modules": "ALL EMBEDDING",
        "lora_rank": 8,
        "lora_alpha": 32
    },
    "experiment": [
      {
        "name": "dpolora",
        "args": {
          "sft_beta": 0.0
        }
      },
      {
        "name": "dpolora+sft_beta0.1",
        "args": {
          "sft_beta": 0.1
        }
      }
    ]
}