conf/Pretrain_v1.5_woInstruction.yaml

# task
task_type: SchemaGuidedInstructBertTask
task_name: Mirror_Pretrain_DataV1.5_woInstruction
comment: '~~content as label, (start, end + 1) span'

# data preprocessing
max_seq_len: 512
debug_mode: false
label_span: tag  # tag `[LM]` or content `person`
mode: span  # w2 (1,2,3) or span (1,3)
stream_mode: false

# filepaths
plm_dir: microsoft/deberta-v3-large
data_dir: resources/Mirror/v1.5/merged/t-rex-200k-woInstruction/remove_instruction
output_dir: mirror_outputs
task_dir: ${output_dir}/${task_name}
train_filepath: ${data_dir}/train.jsonl
dev_filepath: ${data_dir}/dev.jsonl
test_filepath: ${data_dir}/test.jsonl
dump_cache_dir: ${task_dir}/cache
regenerate_cache: false

# training
random_seed: 1227
base_model_path: null
eval_on_data: [train]
select_best_on_data: train
select_best_by_key: loss
final_eval_on_test: false
save_every_ckpt: true
save_best_ckpt: true

warmup_proportion: 0.1
num_epochs: 3
epoch_patience: -1
num_steps: -1
step_patience: -1
step_eval_interval: 10000
train_batch_size: 8
eval_batch_size: 8
grad_accum_steps: 1
learning_rate: !!float 2e-5
other_learning_rate: !!float 1e-4
max_grad_norm: 1.0
weight_decay: 0.1

# model
dropout: 0.3
use_rope: true
biaffine_size: 512