run_args.json

{
  "base_model": "<LOCAL_PATH>/Qwen3-8B",
  "task": "metamath",
  "peft_method": "lora",
  "output_dir": "<LOCAL_PATH>/metamath",
  "dataset_path": "<LOCAL_PATH>/metamathqa-train.parquet",
  "train_profile": "none",
  "force_max_steps": false,
  "max_steps": null,
  "num_train_epochs": 3.0,
  "stop_at_step": null,
  "stop_at_epoch": null,
  "resume_from_checkpoint": null,
  "per_device_train_batch_size": 16,
  "gradient_accumulation_steps": 2,
  "global_train_batch_size": 32,
  "lr": 0.0001,
  "warmup_ratio": 0.05,
  "weight_decay": 0.0,
  "min_lr_ratio": null,
  "lr_scheduler_type": "cosine",
  "adam_beta1": 0.9,
  "adam_beta2": 0.999,
  "grad_clip": 1.0,
  "max_seq_len": 4096,
  "seed": 44,
  "save_strategy": "epoch",
  "save_steps": 500,
  "save_total_limit": 1,
  "logging_steps": 25,
  "max_train_samples": 50000,
  "dataset_seed": 42,
  "bf16": true,
  "fp16": false,
  "gradient_checkpointing": true,
  "group_by_length": false,
  "pad_to_multiple_of": 8,
  "dataloader_num_workers": 0,
  "optim": null,
  "use_qlora": false,
  "sft_format": "chat",
  "chat_template_mode": "non_thinking",
  "target_modules": "all",
  "r": 16,
  "lora_alpha": 32,
  "lora_dropout": 0.05,
  "pissa_init_mode": "pissa",
  "init_r": 32,
  "target_r": 16,
  "loraplus_lr_ratio": 20.0,
  "loraplus_lr_embedding": null
}