TheoremQA / metadata.json
ZTWHHH's picture
metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
7685f2c verified
Raw
History Blame Contribute Delete
2.66 kB
{
"name": "TheoremQA",
"release_date": "2023-05-21",
"subsets": {
"image": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "short_answer_qa",
"score_pipeline": [
"exact-match",
"rule-match"
],
"score_params": {
"numeric_rel_tol": 0.04
},
"score_protocol": {
"reference": "official TIGER-AI-Lab/TheoremQA@utils.py compare_answer_with_groundtruth + number_utils.py compare_two_numbers/within_eps (float: 4% relative tolerance eps=abs(gt)*0.04; int: round(p)==gt; lists element-wise after sorting; option '(a)'..'(f)' substring; strings case-insensitive equality); run_gpt.py splits the response on the literal 'answer is ' — no LLM in extraction or grading. [Re-verified 2026-07-07 against the live repo: eps = abs(gt) * 0.04 in number_utils.py; round(p)==gt for ints; sorted element-wise list compare.]",
"note": "Official inference prompt instructs 'Therefore, the answer is ...' which the rule split relies on; the mm-eval prompt is bare <image>{{question}} with no final-answer instruction, so extraction robustness is reduced (template matchers must handle free-form output). mm-eval ships only the 53-row image subset of the 800-question benchmark (documented in metadata); official overall accuracy not reproducible. List-typed answers (e.g. '[2, 2]') need element-wise numeric comparison, beyond plain exact matching."
},
"prompt_template": "<image>{{ question }}",
"prompt_template_source": {
"origin": "official",
"reference": "https://huggingface.co/datasets/TIGER-Lab/TheoremQA",
"notes": "Tier 1: TheoremQA image subset — multimodal theorem-driven open-ended QA. Filtered to rows with Picture column non-null (text-only rows excluded to fit single-image VQA schema; preserve image subset under \"image\" subset name)."
},
"mapping_from_source": {
"media": {
"from": "images",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "id"
},
"question": {
"from": "question"
},
"answer": {
"from": "answer",
"optional": true
},
"extra": {
"n_images": {
"from": "n_images"
},
"answer_type": {
"from": "answer_type"
}
},
"source": {
"format": "json",
"url": {
"test": "https://huggingface.co/datasets/TIGER-Lab/TheoremQA"
}
}
}
}
}
}