training_args.json

{
  "output_dir": "./results",
  "num_train_epochs": 10,
  "per_device_train_batch_size": 16,
  "gradient_accumulation_steps": 1,
  "learning_rate": 0.0001,
  "weight_decay": 0.01,
  "logging_steps": 100,
  "save_steps": 1000,
  "eval_steps": 500,
  "warmup_steps": 500,
  "max_grad_norm": 1.0,
  "fp16": false,
  "dataloader_num_workers": 4,
  "load_best_model_at_end": true,
  "metric_for_best_model": "loss",
  "greater_is_better": false,
  "report_to": [],
  "seed": 42
}