boolq_bert_large_cased_softprompt.yaml

dataset:
  name: super_glue.boolq
  path:  # dataset in huggingface doesn't need path

plm:
  model_name: bert
  model_path: bert-large-cased
  optimize:
    freeze_para: True
    lr: 1.0e-5
    weight_decay: 0.0
    scheduler:
      type: 
      num_warmup_steps: 500

dataloader:
  max_seq_length: 384 # max_seq_length 
  decoder_max_length: 3 # the decoder max length to truncate decoder input sequence
                    # if it is an encoder-decoder architecture. Note that it's not equavalent
                    # to generation.max_length which is used merely in the generation phase.
  truncate_method: "head" # choosing from balanced, head, tail
  decode_from_pad: false

train:
  batch_size: 4
  gradient_accumulation_steps: 2
  max_grad_norm: 1.0
  num_epochs: 
  num_training_steps: 30000


test:
  batch_size: 16

dev:
  batch_size: 16


template: soft_template
verbalizer: manual_verbalizer



soft_template:
  choice: 0
  file_path: scripts/SuperGLUE/BoolQ/soft_template.txt
  num_tokens: 20
  initialize_from_vocab: true
  random_range: 0.5
  optimize: 
    name: AdamW
    lr: 0.03
    adam_epsilon: 1.0e-8
    scheduler:
      num_warmup_steps: 500


manual_verbalizer:
  choice: 0
  file_path: scripts/SuperGLUE/BoolQ/manual_verbalizer.txt
  
environment:
  num_gpus: 1
  cuda_visible_devices:
  local_rank: 0 

learning_setting: full #few_shot

# few_shot:
#   parent_config: learning_setting
#   few_shot_sampling: sampling_from_train
  
# sampling_from_train:
#   parent_config: few_shot_sampling
#   num_examples_per_label: 100
#   also_sample_dev: True
#   num_examples_per_label_dev: 100
#   seed:
#     - 123