sft_config.json

{
  "sft_experiment_id": "E-9",
  "dataset_strategy_id": "baseline",
  "base_model": "unsloth/Qwen3-4B-Instruct-2507",
  "neftune_alpha": 5.0,
  "data_cleaning": true,
  "lora": {
    "r": 64,
    "alpha": 128,
    "dropout": 0.0,
    "target_modules": [
      "q_proj",
      "k_proj",
      "v_proj",
      "o_proj",
      "gate_proj",
      "up_proj",
      "down_proj"
    ]
  },
  "training": {
    "max_seq_len": 512,
    "epochs": 2,
    "lr": 2e-06,
    "per_device_train_bs": 2,
    "grad_accum": 8,
    "warmup_ratio": 0.1,
    "weight_decay": 0.05
  },
  "special": {
    "mask_cot": true,
    "upsampling": true,
    "curriculum": false
  }
}