# task
task_type: MrcTaggingTask
task_name: debug-Mirror_W2_MSRAv2_NER_FreezeBertEmbAnd0-3_bs64
comment: 'bert mrc w/ w2ner for NER'

# data preprocessing
max_seq_len: 300
negative_sample_prob: 1.0
debug_mode: false
mode: w2

# filepaths
base_model_path: outputs/RobertaBase_data20230314v2/ckpt/MrcGlobalPointerModel.best.pth
plm_dir: hfl/chinese-roberta-wwm-ext
data_dir: resources/NER/MSRA_v2/formatted
output_dir: outputs
task_dir: ${output_dir}/${task_name}
train_filepath: ${data_dir}/train.char.bmes.jsonl
dev_filepath: ${data_dir}/dev.char.bmes.jsonl
test_filepath: ${data_dir}/test.char.bmes.jsonl
ent_type2query_filepath: ${data_dir}/query.json
dump_cache_dir: ${task_dir}/cache
regenerate_cache: true

# training
random_seed: 1227
eval_on_data: [dev, test]
select_best_on_data: dev
select_best_by_key: metric
best_metric_field: micro.f1
final_eval_on_test: true

warmup_proportion: 0.1
num_epochs: 5
epoch_patience: 5
train_batch_size: 64
eval_batch_size: 128
learning_rate: !!float 5e-5
other_learning_rate: !!float 1e-4
max_grad_norm: 1.0
weight_decay: 0.1

# model
dropout: 0.3
biaffine_size: 512