MMVet / metadata.json
ZTWHHH's picture
metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
562fd22 verified
Raw
History Blame Contribute Delete
2.47 kB
{
"name": "MMVet",
"release_date": "2026-05-15",
"subsets": {
"main": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "short_answer_qa",
"score_pipeline": [
"llm-judge"
],
"score_params": {
"rubric": "partial_credit",
"scale": [
0.0,
1.0
],
"official_judge_model": "gpt-4-0613",
"reference_syntax": "and_or_composition"
},
"score_protocol": {
"reference": "VLMEvalKit@vlmeval/dataset/utils/mmvet.py:8-71 — GPT-4 few-shot prompt assigns per-sample correctness 0.0..1.0 (0.1 steps) against gold with <AND>/<OR> multi-component semantics; vlmeval/dataset/image_vqa.py:2037-2098 (MMVet.evaluate). Same prompt in lmms-eval@lmms_eval/tasks/mmvet/utils.py (MM_VET_PROMPT). Official: yuweihao/MM-Vet mm-vet_evaluator (gpt-4-0613).",
"note": "Official headline = mean per-sample judge score x100 (Overall + per-capability rec/ocr/know/gen/spat/math breakdown via `capability` extra). Per-sample score is fractional partial credit, not binary; VLMEvalKit falls back to score 0.0 after 5 failed judge parses. lmms-eval currently defaults the judge to gpt-4o-2024-11-20 while the official evaluator used gpt-4-0613."
},
"prompt_template": "<image>{{ question }}",
"mapping_from_source": {
"media": {
"from": "image",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "id"
},
"question": {
"from": "question"
},
"answer": {
"from": "answer",
"optional": true,
"note": "uses `<OR>` / `<AND>` separators for alternative gold answers"
},
"extra": {
"capability": {
"from": "capability"
}
},
"source": {
"format": "json",
"url": {
"test": "https://huggingface.co/datasets/whyu/mm-vet"
}
}
},
"prompt_template_source": {
"origin": "official",
"reference": "https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/dataset/image_vqa.py (MMVet — judge-scored; bare question with no harness suffix)",
"notes": "Tier 3: MMVet judge-scored eval: bare question (no harness suffix) per the official MMVet paper and VLMEvalKit handling."
}
}
}
}