File size: 2,231 Bytes
5d11daa 0708eb2 5d11daa 0708eb2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 | {
"name": "MME",
"release_date": "2023-06-23",
"subsets": {
"main": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "yes_no_qa",
"score_pipeline": [
"exact-match",
"rule-match"
],
"score_protocol": {
"reference": "lmms-eval@lmms_eval/tasks/mme/utils.py:48-90 — parse_pred_ans rule (yes/no/y/n/prefix match, else 'other') taken from the official MME calculation script; utils.py:93-123 official score = acc + acc+ (acc+ requires both questions of an image correct), summed over categories.",
"note": "Official aggregation not reproduced per-sample: per-category score = acc + acc+ where acc+ needs pairwise grouping of the 2 questions per image (pair key = extra.source_id, ids <img>_q1/<img>_q2); perception total (max 2000) and cognition total (max 800) are sums over categories. Per-sample scoring yields plain accuracy only. VLMEvalKit variant adds GPT yes/no extraction fallback (image_yorn.py:52-85 + yorn.py:264-272 YOrN_auxeval) before MME_rating (yorn.py:50-93); official/lmms-eval protocol is pure rule — rule chosen per official."
},
"prompt_template": "<image>{{ question }}",
"mapping_from_source": {
"media": {
"from": "image",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "question_id"
},
"question": {
"from": "question"
},
"answer": {
"from": "answer",
"optional": true
},
"extra": {
"category": {
"from": "category"
}
},
"source": {
"format": "huggingface",
"url": {
"test": "https://huggingface.co/datasets/lmms-lab/MME"
}
}
},
"prompt_template_source": {
"origin": "official",
"reference": "https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/lmms_eval/tasks/mme/utils.py (mme_doc_to_text — bare question; answer is yes/no)",
"notes": "Tier 4: lmms-eval MME canonical evaluation prompt (perception/cognition yes/no eval)."
}
}
}
} |