train_config.json

{
  "train_dataset": "/workspace/fms4navigation/datasets/Faithfulness-Critic-Dataset/train_dataset.jsonl",
  "val_dataset": "/workspace/fms4navigation/datasets/Faithfulness-Critic-Dataset/val_dataset_512.jsonl",
  "dataset_dir": "/workspace/fms4navigation/datasets/Faithfulness-Critic-Dataset",
  "max_length": 4096,
  "seed": 0,
  "eval_max_new_tokens": 96,
  "eval_steps": 100,
  "balance_by_inconsistent_count": false,
  "balance_alpha": 1.0,
  "base_model": "Qwen/Qwen3-VL-4B-Instruct",
  "use_lora": true,
  "lora_rank": 256,
  "lora_alpha": null,
  "lora_dropout": 0.05,
  "lora_target_modules": "q_proj,k_proj,v_proj,o_proj",
  "flash_attn": false,
  "gradient_checkpointing": false,
  "output_dir": "/workspace/fms4navigation/results/PRM-v2-r256",
  "epochs": 1.0,
  "per_device_batch_size": 4,
  "grad_accum_steps": 1,
  "lr": 0.0001,
  "warmup_ratio": 0.03,
  "weight_decay": 0.0,
  "logging_steps": 20,
  "save_steps": 300,
  "save_total_limit": 3,
  "max_steps": -1,
  "report_to": "wandb"
}