config.yaml

trainer:
  device: &device cuda
  random_seed: 42
  batch_size_no_grad: 64
  batch_size_grad: 32
  max_epoch: 16
  weight_only_epoch: 0
  fix_weights: false
  patience: 4

  leave_out: false
  force_single_token: true
  shuffle: true
  show_top_patterns: 0
  training_target:
    - conditional
    # - generate
  accumulate_gradient: 1
  randomize_prompt: false
  frequent: false
  smoothing: 0.0

  max_layer: 24
  penalty: [0.0, 0.0, 0.0, 0.0, 0.0]
  log_path: 'logs/trex'

  vocab_file: &vocab_file ~
  conditional_prompt: false

db:
  path: db/trex_extend
  single_lexicon: true
  relation_type_filter: &filter
  vocab_file: *vocab_file

template:
  relation_type_filter: *filter
  path: prompts/trex_mine
  min_length: -1
  max_length: -1
  max_num: -1

lm:
  model_type: 'bert'
  param_name: 'bert-large-cased'
  device: *device