TrainConfig:
  epochs: 6
  eval_iter: 2000
  learning_rate: 3.0e-5
  optimizer_builder:
    optimizer: 
      type: AdamW
    weight_decay: 0.01
  origin_metric: 0.8098