bert.json

{
  "model_name": "ru_bert_parser_with_tagger",
  "embedder": {
    "name": "ru_bert",
    "dropout": 0.4,
    "use_pymorphy": true
  },
  "encoder": {
    "encoder_type": "lstm",
    "hidden_dim": 256,
    "num_layers": 1,
    "dropout": 0.3,
    "variational_dropout": 0,
    "use_weight_drop": false
  },
  "parser": {
    "dropout": 0.1,
    "tag_representation_dim": 256,
    "arc_representation_dim": 512,
    "gram_val_representation_dim": 256,
    "lemma_representation_dim": 256
  },
  "trainer": {
    "batch_size": 64,
    "num_epochs": 25,
    "patience": 5,
    "lr": 1e-3,
    "bert_lr": 1e-4,
    "cut_frac": 0.1,
    "gradual_unfreezing": true,
    "discriminative_fine_tuning": true,
    "num_gradient_accumulation_steps": 1
  },
  "data": {
    "data_dir": "../data",
    "pretrained_models_dir": "../pretrained_models",
    "models_dir": "../models",
    "train_data_all_except": null,
    "train_data": [
      "GramEval2020-17cent-train.conllu",
      "GramEval2020-GSD-train.conllu",
      "GramEval2020-SynTagRus-train-v2.conllu",
      "GramEval2020-Taiga-news-train.conllu",
      "GramEval2020-Taiga-poetry-train.conllu",
      "GramEval2020-Taiga-social-train.conllu",
      "MorphoRuEval2017-JZ-gold.conllu",
      "MorphoRuEval2017-Lenta-train.conllu",
      "MorphoRuEval2017-VK-gold.conllu"
    ],
    "valid_data": [
      "GramEval2020-17cent-dev.conllu",
      "GramEval2020-GSD-wiki-dev.conllu",
      "GramEval2020-RuEval2017-Lenta-news-dev.conllu",
      "GramEval2020-RuEval2017-social-dev.conllu",
      "GramEval2020-SynTagRus-dev.conllu",
      "GramEval2020-Taiga-poetry-dev.conllu"
    ]
  }
}