mgsm_chat.yaml

tag: llama3
task: mgsm_chat
dataset_path: meta-llama/Llama-3.2-3B-Instruct-evals
dataset_name: Llama-3.2-3B-Instruct-evals__mgsm__details
output_type: generate_until
test_split: latest
doc_to_text: "{{
    input_final_prompts
    |first
    |replace('<|start_header_id|>user<|end_header_id|>', '')
    |replace('<|eot_id|><|start_header_id|>assistant<|end_header_id|>', '')
    |trim
}}"
doc_to_target: "input_correct_responses"
process_results: !function utils.process_results_mgsm
generation_kwargs:
  until: []
  do_sample: false
  temperature: 0.0
  max_gen_toks: 2048
metric_list:
  - metric: exact_match
    aggregation: mean
    higher_is_better: true
    ignore_case: true
    ignore_punctuation: true
filter_list:
  - name: "strict-match"
    filter:
      - function: "regex"
        regex_pattern: "(?:Answer|Réponse|Antwort|Ответ|Respuesta|答え|Jibu|答案|คำตอบ|సమాధానం|উত্তর): (\\-?[0-9\\.\\,]+)"
      - function: remove_whitespace
      - function: take_first
  - name: "flexible-extract"
    filter:
      - function: regex
        group_select: -1
        regex_pattern: "(?:Answer|Réponse|Antwort|Ответ|Respuesta|答え|Jibu|答案|คำตอบ|సమాధానం|উত্তর): (-?[$0-9.,]{2,})|(-?[0-9]+)"
      - function: remove_whitespace
      - function: take_first
metadata:
  version: 0.0