trainer_state.json

{
  "global_step": 125,
  "step": 1000,
  "best_eval_hit_rate": -1.0,
  "last_metrics": {},
  "config": {
    "model_name": "mistralai/Ministral-3-8B-Instruct-2512-BF16",
    "learning_rate": 5e-06,
    "lora_r": 16,
    "grpo_beta": 0.0,
    "num_generations": 4,
    "max_steps": -1,
    "output_dir": "./outputs/ministral_duckhunt_grpo"
  }
}