MMEvalPro / metadata.json
ZTWHHH's picture
metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
2ad19dc verified
Raw
History Blame Contribute Delete
3.25 kB
{
"name": "MMEvalPro",
"release_date": "2026-07-07",
"subsets": {
"main": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "multiple_choice_qa",
"score_pipeline": [
"exact-match",
"rule-match"
],
"score_protocol": {
"reference": "official@github.com/chenllliang/MMEvalPro auto_score.py — no extraction/LLM: direct equality answer == model_output per question; Genuine Accuracy groups rows by triplet_id and counts a triplet correct only if all three (Origin/Perception/Knowledge) are correct; also Average Accuracy and per-eval_type accuracies, macro/micro over source datasets.",
"note": "Official HEADLINE metric is Genuine Accuracy over question triplets: group rows by 'triplet_id' (each triplet has exactly 3 rows, one per eval_type Origin/Perception/Knowledge) and count a triplet correct only if all 3 of its rows are answered correctly; Genuine Accuracy = correct_triplets / total_triplets (2138 triplets). The per-sample scorer computes only the official Average Accuracy (per-row equality answer==prediction); reconstruct Genuine Accuracy downstream by grouping result rows on the carried 'triplet_id' field. Source: github.com/chenllliang/MMEvalPro auto_score.py (triplet_id + eval_type grouping)."
},
"prompt_template": "<image>Question: {{ question }}\nOptions:\n{% for k, v in options.items() %}{{ k }}. {{ v }}\n{% endfor %}Please select the correct answer from the options above. ",
"prompt_template_source": {
"origin": "official",
"reference": "VLMEvalKit vlmeval/dataset/image_mcq.py (ImageMCQDataset.build_prompt — deployed rendering)",
"notes": "Tier 3 deployed. The authors' own prompt exists (arXiv:2407.00468 App. B.1: 'Analyse the image and choose the best answer...Just output the letter of the correct answer.') but its {options} byte-format is under-specified in the paper — flagged as an open question rather than deploying an invented rendering. The fabricated '{% if hint %}' branch was removed 2026-07-07 (source schema has no hint field; rendering unchanged since hint was always empty)."
},
"mapping_from_source": {
"media": {
"from": "image",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "index"
},
"question": {
"from": "question"
},
"answer": {
"from": "answer",
"optional": true
},
"options": {
"from": "choices",
"optional": true,
"note": "list source values are normalized to {A,B,...} dict"
},
"extra": {
"source": {
"from": "source"
},
"question_category": {
"from": "question_category"
},
"eval_type": {
"from": "eval_type"
},
"triplet_id": {
"from": "triplet_id"
}
},
"source": {
"format": "huggingface",
"url": {
"test": "https://huggingface.co/datasets/MM-Diagnose/MMEvalPro"
}
}
}
}
}
}