train_args.json

{
  "output_dir": "./output/pum_qwen3_4b_lora_newdata_0420",
  "per_device_train_batch_size": 8,
  "per_device_eval_batch_size": 8,
  "gradient_accumulation_steps": 4,
  "num_train_epochs": 2,
  "learning_rate": 1e-05,
  "weight_decay": 0.01,
  "warmup_ratio": 0.08,
  "max_grad_norm": 1.0,
  "logging_steps": 5,
  "eval_steps": 20,
  "save_steps": 200,
  "bf16": true,
  "fp16": false,
  "gradient_checkpointing": true,
  "gradient_checkpointing_use_reentrant": false,
  "ddp_static_graph_for_lora": true,
  "num_workers": 4,
  "seed": 42,
  "log_jsonl_name": "training_log.jsonl",
  "save_png_curves": true,
  "plot_dpi": 160,
  "tie_eval_margin": 0.25,
  "train_random_swap_prob": 0.0,
  "eval_random_swap_prob": 0.0
}