config.json

{
  "model_name": "Qwen/Qwen3-1.7B",
  "init_checkpoint": "",
  "ref_model_name": "",
  "skip_first_n_prompts": 0,
  "dataset_name": "cais/mmlu",
  "mmlu_config": "all",
  "mmlu_split": "test",
  "lambda_length": 0.0,
  "format_penalty": -1.0,
  "max_steps": 250,
  "per_device_train_batch_size": 0,
  "gradient_accumulation_steps": 6,
  "learning_rate": 5e-05,
  "warmup_ratio": 0.05,
  "lr_scheduler_type": "constant_with_warmup",
  "max_grad_norm": 1.0,
  "bf16": true,
  "gradient_checkpointing": true,
  "num_generations": 6,
  "max_completion_length": 3000,
  "max_prompt_length": 512,
  "temperature": 0.6,
  "top_p": 0.95,
  "loss_type": "dr_grpo",
  "beta": 0.0,
  "num_iterations": 1,
  "epsilon": 0.2,
  "scale_rewards": "group",
  "lora_rank": 16,
  "lora_alpha": 32,
  "lora_dropout": 0.05,
  "lora_target_modules": [
    "q_proj",
    "k_proj",
    "v_proj",
    "o_proj",
    "gate_proj",
    "up_proj",
    "down_proj"
  ],
  "use_vllm": true,
  "vllm_mode": "colocate",
  "vllm_gpu_memory_utilization": 0.2,
  "vllm_model_impl": "transformers",
  "logging_steps": 1,
  "log_completions": false,
  "report_to": "wandb",
  "save_steps": 50,
  "save_total_limit": 5,
  "output_dir": "./outputs",
  "run_name": "hint_follow_lr5e5_250steps",
  "seed": 42,
  "baseline_num_problems": 100,
  "baseline_generations_per_problem": 8,
  "wandb_project": "grpo-mmlu-hint",
  "lora_config": {
    "r": 16,
    "lora_alpha": 32,
    "lora_dropout": 0.05,
    "target_modules": [
      "v_proj",
      "o_proj",
      "q_proj",
      "gate_proj",
      "k_proj",
      "down_proj",
      "up_proj"
    ],
    "bias": "none",
    "task_type": "TaskType.CAUSAL_LM"
  }
}