TrainConfig: epochs: 16 eval_iter: 1000 learning_rate: 1.0e-5
optimizer_builder: optimizer: type: AdamW
weight_decay: 0.01 origin_metric: 0.7736
