{
  "per_device_train_batch_size": 4,
  "per_device_eval_batch_size": 4,
  "gradient_accumulation_steps": 8,
  "learning_rate": 0.0003,
  "num_train_epochs": 15,
  "max_steps": -1
}