training_config.json

{
  "num_train_epochs": 3,
  "per_device_train_batch_size": 4,
  "gradient_accumulation_steps": 4,
  "learning_rate": 0.0002,
  "warmup_steps": 100,
  "fp16": true,
  "dataset_size": 8086,
  "blocked_portuguese": 13390
}