{
  "experiment_name": "caost-test-2",
  "trainer": "sft",
  "save_model": {
    "type": "git",
    "model_repository": "git@git.dev2.aifrica.co.kr:groupuser/gemma-3-1b-it-finetuning.git"
  },
  "dataset": {
    "type": "git",
    "dataset_branch": "main",
    "dataset_repository": "git@git.dev2.aifrica.co.kr:groupuser/BCCard-Finance-Kor-QnA-Small.git",
    "dataset_path": "/cheetah/input/dataset/groupuser/BCCard-Finance-Kor-QnA-Small"
  },
  "tokenizer_parameters": {
    "block_size": 1024,
    "max_length": 1024,
    "padding": "right",
    "add_eos_token": true
  },
  "model": {
    "type": "git",
    "model_repository": "git@git.dev2.aifrica.co.kr:groupuser/gemma-3-4b-it.git",
    "model_branch": "main",
    "model_path": "/cheetah/input/model/groupuser/gemma-3-4b-it"
  },
  "peft": {
    "target_modules": "all-linear",
    "quantization": "int4",
    "merge_adapter": null,
    "tuner": "lora",
    "r": 16,
    "lora_alpha": 32,
    "lora_dropout": 0.05,
    "bias": "none",
    "task_type": "CAUSAL_LM",
    "peft": true
  },
  "train_model_parameters": {
    "log": "tensorboard",
    "scheduler": "linear",
    "seed": 42,
    "batch_size": 2,
    "output_dir": "/cheetah/fine-tuning/output",
    "optimizer": "adamw_torch",
    "dataset_text_field": "text",
    "column_mappings": {
      "prompt_text_column": "",
      "rejected_text_column": "text",
      "text_column": "text"
    },
    "logging_steps": -1,
    "logging_strategy": "epoch",
    "use_flash_attention": false,
    "evaluation_strategy": "epoch",
    "save_total_limit": 1,
    "auto_find_batch_size": true,
    "mixed_precision": "fp16",
    "learning_rate": 3e-05,
    "chat_template": "None",
    "max_prompt_length": null,
    "max_completion_length": null,
    "distributed_backend": "None",
    "num_train_epochs": 1,
    "warmup_ratio": 0.1,
    "weight_decay": 0,
    "max_grad_norm": 1,
    "model_ref": "",
    "dpo_beta": 0.1,
    "use_fsdp2": false,
    "disable_gc": false,
    "unsloth": "false",
    "do_train": true,
    "do_predict": true,
    "gradient_checkpointing": true,
    "per_device_train_batch_size": 4,
    "per_device_eval_batch_size": 4,
    "gradient_accumulation": 4
  }
}