dataset_path: kcz358/DynaMath
output_type: generate_until
doc_to_target: "ground_truth"
metric_list:
  - metric: dynamath_average
    aggregation: !function utils.dynamath_aggregate_results_average
    higher_is_better: true
  - metric: dynamath_worst
    aggregation: !function utils.dynamath_aggregate_results_worst
    higher_is_better: true

generation_kwargs:
  max_new_tokens: 49152

test_split: test
doc_to_messages: !function utils.dynamath_doc_to_messages_cot
doc_to_text: !function utils.dynamath_doc_to_text_cot
doc_to_visual: !function utils.dynamath_doc_to_visual
process_results: !function utils.dynamath_process_results