# @package _group_
common:
  fp16: true
  log_format: json
  log_interval: 200

checkpoint:
  no_epoch_checkpoints: true
  save_interval: 5
  save_interval_updates: 50000
  
task:
  _name: masked_lm
  data: ???
  sample_break_mode: complete
  tokens_per_sample: 512

criterion: masked_lm



optimizer:
  _name: adan
  weight_decay: 0.02
  adan_betas: (0.98,0.92,0.99)
  adan_eps: 1e-08

lr_scheduler:
  _name: polynomial_decay
  warmup_updates: 10000

optimization:
  clip_norm: 5.0
  lr: [0.001]
  max_update: 1000000
  update_freq: [1]

model:
  _name: roberta
  max_positions: 512
  dropout: 0.1
  attention_dropout: 0.1

distributed_training:
  ddp_backend: no_c10d

dataset:
  skip_invalid_size_inputs_valid_test: true
  validate_interval: 5
  validate_interval_updates: 50000
  batch_size: 32
  ignore_unused_valid_subsets: true