# task
task_type: MrcQaTask
task_name: Mirror_RobertaBaseWwm_Cons_MsraMrc
comment: 'GlobalPointer with RoPE'

# data preprocessing
max_seq_len: 512
debug_mode: false
mode: cons

# filepaths
plm_dir: hfl/chinese-roberta-wwm-ext
data_dir: resources/NER/msra/mrc
output_dir: outputs
task_dir: ${output_dir}/${task_name}
train_filepath: ${data_dir}/train.jsonl
dev_filepath: ${data_dir}/test.jsonl
test_filepath: ${data_dir}/test.jsonl
dump_cache_dir: ${task_dir}/cache
regenerate_cache: true

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

warmup_proportion: 0.1
step_eval_interval: 20000
step_patience: -1
num_epochs: 5
epoch_patience: 5
train_batch_size: 32
eval_batch_size: 64
learning_rate: !!float 5e-5
other_learning_rate: !!float 1e-4
max_grad_norm: 1.0

# model
dropout: 0.3
biaffine_size: 512