train_summary.json

{
  "base_model": "Qwen/Qwen3-4B-Instruct-2507",
  "data_dir": "data/dataset_v2",
  "train_examples": 1288,
  "epochs": 1.0,
  "max_steps": 800,
  "optimizer_steps": 800,
  "per_device_batch_size": 1,
  "grad_accum": 8,
  "max_seq_len": 1024,
  "learning_rate": 0.0002,
  "lora_r": 32,
  "lora_alpha": 64,
  "lora_dropout": 0.05,
  "seed": 0,
  "qlora": true,
  "device": "cuda",
  "lr_scheduler_type": "cosine",
  "warmup_ratio": 0.03,
  "weight_decay": 0.001,
  "max_grad_norm": 0.3,
  "optim": "paged_adamw_32bit",
  "final_train_loss": 0.3792
}