MMVU / metadata.json
ZTWHHH's picture
metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
eb397ba verified
Raw
History Blame Contribute Delete
2.93 kB
{
"name": "MMVU",
"release_date": "2025-01-01",
"subsets": {
"main": {
"language": [
"en"
],
"modalities": [
"single_video_start"
],
"task_type": "multiple_choice_qa",
"score_pipeline": [
"exact-match",
"rule-match",
"llm-judge"
],
"score_protocol": {
"reference": "official@github.com/yale-nlp/MMVU acc_evaluation.py + utils/eval_utils.py — GPT-4o extracts the final answer and outputs a binary correctness flag for BOTH multiple-choice and open-ended questions; lmms-eval@lmms_eval/tasks/mmvu/utils.py:144-300 — hybrid: deterministic rule matching first (letter/substring/numeric with 0.01 tolerance), GPT judge (binary, gpt-4o default) fallback for open-ended only.",
"note": "Official (yale-nlp) grades every sample with GPT-4o; the rule-first/judge-fallback execution matches lmms-eval's hybrid and is the endorsed efficiency deviation. Subset mixes MCQ (5 options A-E) and open-ended rows; extra.question_type carries source values 'multiple-choice'/'open-ended' which are NOT in the scorer vocabulary. Official judge model gpt-4o."
},
"prompt_template": "<video>{{ question }}{% if options %}\n{% for k, v in options.items() %}{{ k }}. {{ v }}\n{% endfor %}Answer with the option's letter from the given choices directly.{% endif %}",
"mapping_from_source": {
"media": {
"from": "video_path",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "id"
},
"question": {
"from": "question"
},
"answer": {
"from": "answer",
"optional": true
},
"options": {
"from": "options",
"optional": true,
"note": "list source values are normalized to {A,B,...} dict"
},
"extra": {
"video_id": {
"from": "video_id"
},
"question_type": {
"from": "question_type"
},
"subject": {
"from": "subject"
}
},
"source": {
"format": "json",
"url": {
"validation": "https://huggingface.co/datasets/yale-nlp/MMVU"
}
}
},
"video_storage": {
"format": "files",
"media_root": "videos",
"notes": "583 videos from the MMVU benchmark (yale-nlp/MMVU), hosted on HF. Covers 27 expert disciplines. Videos sourced from YouTube."
},
"prompt_template_source": {
"origin": "official",
"reference": "https://github.com/yale-nlp/MMVU/blob/main/utils/constant.py (MULTI_CHOICE_DO_PROMPT / OPEN_ENDED_DO_PROMPT — yale-nlp MMVU direct-output eval prompts)",
"notes": "Tier 1: MMVU official evaluation prompts (yale-nlp MMVU paper authors): MCQ + open-ended branches."
}
}
}
}