run_config.json

{
  "base_model": "<LOCAL_PATH>/Qwen3-8B",
  "task": "magicoder",
  "peft_method": "lora",
  "train_profile": "none",
  "num_train_epochs": 1.0,
  "max_steps": null,
  "stop_at_step": null,
  "stop_at_epoch": null,
  "resume_from_checkpoint": null,
  "training_budget": "epochs",
  "per_device_train_batch_size": 16,
  "gradient_accumulation_steps": 2,
  "global_train_batch_size": 32,
  "world_size": 1,
  "effective_global_batch_size": 32,
  "steps_per_epoch": 1563,
  "total_steps": 1563,
  "lr": 2e-05,
  "warmup_ratio": 0.05,
  "weight_decay": 0.0,
  "lr_scheduler_type": "cosine",
  "min_lr_ratio": null,
  "max_seq_len": 4096,
  "dataset_size": 50000,
  "dataset_path": "<LOCAL_PATH>/magicoder-train.parquet",
  "max_train_samples": 50000,
  "dataset_seed": 42,
  "sft_format": "chat",
  "chat_template_mode": "non_thinking",
  "save_strategy": "epoch",
  "save_steps": 500,
  "save_total_limit": 1,
  "logging_steps": 25,
  "adalora_schedule": null
}