group: bbh_fewshot
dataset_path: lukaemon/bbh
output_type: generate_until
test_split: test
doc_to_target: "{{target}}"
filter_list:
  - name: "strict-match"
    filter:
      - function: "take_first"
  - name: "flexible-extract"
    filter:
      - function: !function utils.MultiChoiceRegexFilter
        group_select: 0
        # regex_pattern: "^\\s*(.*?)(?=(\\n+|$))"
        regex_pattern: "^\\s*(.+?)(?=(\\n+|$))"
        ignore_case: true
        ignore_punctuation: true
      - function: "take_first"
metric_list:
  - metric: exact_match
    aggregation: mean
    higher_is_better: true
    # ignore_case: true
    # ignore_punctuation: true
generation_kwargs:
  until:
    - "</s>"
    - "Q"
    - "\n\n"
  do_sample: false
  temperature: 0.0
  max_gen_toks: 12
num_fewshot: 0
metadata:
  version: 1.0
  num_fewshot: 3 # will be printed in results table
