| { |
| "name": "MM-Vet-v2", |
| "release_date": "2024-07-23", |
| "subsets": { |
| "default": { |
| "language": [ |
| "en" |
| ], |
| "modalities": [ |
| "multi_image_start" |
| ], |
| "task_type": "open_ended_qa", |
| "score_type": "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": "lmms-eval@lmms_eval/tasks/mmvetv2/utils.py:193-245 + mmvetv2.yaml — identical MM-Vet GPT-4 few-shot 0.0..1.0 partial-credit prompt with <AND>/<OR>, gpt_eval_score aggregation; official dataset whyu/mm-vet-v2 (Yu et al. 2024) uses the MM-Vet evaluator. Verified in local clone (MM_VET_PROMPT at utils.py:193; yaml gpt_eval_score line 18).", |
| "note": "Official headline = mean judge score x100 overall + per-capability (adds new 'seq' capability in v2). DUPLICATE UPLOAD: mm-eval/MM-Vet-v2 and mm-eval/MMVet-v2 both contain the identical 517-row MM-Vet v2 test set (same ids v2_*, same questions/answers); this copy is the older conversion vintage (modalities multi_image_start, extras reference_response/judge_type duplicating answer) — recommend deprecating one repo after migration." |
| }, |
| "prompt_template": "{{ question }}", |
| "prompt_template_source": { |
| "origin": "official", |
| "reference": "https://huggingface.co/datasets/whyu/mm-vet-v2 (paper: MM-Vet-v2, Yu et al. 2024)", |
| "notes": "Tier 1: MM-Vet-v2 indexed multi-image with <IMG><image_N> markers. <IMG> stripped (image-follows marker); each <image_N> replaced with canonical <image> interleaved at its original position (preserves MM-Vet question layout). Reference uses <AND>/<OR> separators for multi-component gold sets; judge is GPT-4 per MM-Vet eval code." |
| }, |
| "mapping_from_source": { |
| "media": { |
| "from": "images", |
| "type": "list", |
| "min_items": 1, |
| "max_items": 18 |
| }, |
| "id": { |
| "from": "id" |
| }, |
| "question": { |
| "from": "question" |
| }, |
| "answer": { |
| "from": "answer", |
| "optional": true |
| }, |
| "extra": { |
| "n_images": { |
| "from": "n_images" |
| }, |
| "reference_response": { |
| "from": "reference_response" |
| }, |
| "capability": { |
| "from": "capability" |
| }, |
| "added_in": { |
| "from": "added_in" |
| }, |
| "judge_type": { |
| "from": "judge_type" |
| } |
| }, |
| "source": { |
| "format": "json", |
| "url": { |
| "test": "https://huggingface.co/datasets/whyu/mm-vet-v2" |
| } |
| } |
| } |
| } |
| }, |
| "deprecated": { |
| "by": "mm-eval/MMVet-v2", |
| "reason": "duplicate upload of MM-Vet v2 (517 identical rows); MMVet-v2 is the canonical copy per owner decision", |
| "date": "2026-07-07" |
| } |
| } |
|
|
