dataset_path: FlagEval/ERQA
test_split: test
output_type: generate_until
doc_to_visual: !function utils.erqa_doc_to_visual
doc_to_text: !function utils.erqa_doc_to_text
doc_to_messages: !function utils.erqa_doc_to_messages
doc_to_target: "answer"

process_results: !function utils.erqa_process_results

metric_list:
  - metric: erqa_acc
    aggregation: !function utils.erqa_aggregate_results
    higher_is_better: true

generation_kwargs:
  max_new_tokens: 1024