train_args.json

{
    "data_train": [
        "data/nkjp-nested-ttt/train.txt",
        "data/nkjp-nested-ttt/valid.txt",
        "data/nkjp-nested-ttt/test.txt"
    ],
    "data_tune": [
        "data/nkjp-nested-ttt/valid.txt"
    ],
    "data_test": [
        "data/nkjp-nested-ttt/test.txt"
    ],
    "pretrained_path": "allegro/herbert-large-cased",
    "output_dir": "../poldeepner2_models/dev/nkjp_full/model_nkjp_full_union_256_101_v_003",
    "cache_dir": "",
    "device": "cuda:0",
    "max_seq_length": 256,
    "do_eval": false,
    "do_lower_case": false,
    "train_batch_size": 16,
    "eval_batch_size": 16,
    "learning_rate": 5e-06,
    "num_train_epochs": 20,
    "warmup_proportion": 0.0,
    "weight_decay": 0.01,
    "adam_epsilon": 1e-08,
    "max_grad_norm": 1.0,
    "seed": 101,
    "gradient_accumulation_steps": 1,
    "fp16": false,
    "fp16_opt_level": "O1",
    "loss_scale": 0,
    "dropout": 0.2,
    "freeze_model": false,
    "epoch_save_model": true,
    "sequence_generator": "union",
    "sequence_generator_for_eval": "context-window",
    "training_mix": false,
    "wandb": "nkjp_full",
    "hidden_size": 1024
}