{
  "dataset_config": {
    "epochs": 20,
    "batch_size": 192,
    "pre_train_dataset": "",
    "fine_tune_dataset": "",
    "test_dataset": "",
    "valid_dataset": "",
    "dataset_sink_mode": false,
    "dataset_sink_step": 100
  },
  "model_config": {
    "random_seed": 100,
    "save_graphs": false,
    "seq_length": 64,
    "vocab_size": 45744,
    "hidden_size": 1024,
    "num_hidden_layers": 6,
    "num_attention_heads": 8,
    "intermediate_size": 4096,
    "hidden_act": "relu",
    "hidden_dropout_prob": 0.2,
    "attention_dropout_prob": 0.2,
    "max_position_embeddings": 64,
    "initializer_range": 0.02,
    "label_smoothing": 0.1,
    "beam_width": 4,
    "length_penalty_weight": 1.0,
    "max_decode_length": 64,
    "input_mask_from_dataset": true
  },
  "loss_scale_config": {
    "loss_scale_mode": "dynamic",
    "init_loss_scale": 65536,
    "loss_scale_factor": 2,
    "scale_window": 200
  },
  "learn_rate_config": {
    "optimizer": "adam",
    "lr": 1e-4,
    "lr_scheduler": "poly",
    "poly_lr_scheduler_power": 0.5,
    "decay_steps": 10000,
    "decay_start_step": 12000,
    "warmup_steps": 4000,
    "min_lr": 1e-6
  },
  "checkpoint_options": {
    "existed_ckpt": "",
    "save_ckpt_steps": 2500,
    "keep_ckpt_max": 50,
    "ckpt_prefix": "ckpt",
    "ckpt_path": "checkpoints"
  }
}