| { | |
| "name": "MuirBench", | |
| "release_date": "2024-06-06", | |
| "subsets": { | |
| "main": { | |
| "language": [ | |
| "en" | |
| ], | |
| "modalities": [ | |
| "multi_image_interleave" | |
| ], | |
| "task_type": "multiple_choice_qa", | |
| "score_pipeline": [ | |
| "exact-match", | |
| "rule-match" | |
| ], | |
| "score_protocol": { | |
| "reference": "official@github.com/muirbench/MuirBench eval/utils/postprocess.py:13 (parse_multi_choice_response) — deterministic string matching: '(A)' form, then isolated ' A ', then option-content substring for long responses; random option as last resort (L42-43, random.seed(42)); no LLM. lmms-eval@lmms_eval/tasks/muirbench/muirbench.yaml:23-37 + utils.py (MultiChoiceRegexFilter + lowercase exact match, muir_aggregation) agrees.", | |
| "note": "Official falls back to a RANDOM choice when nothing matches — a per-sample scorer should instead mark unparseable responses incorrect (deterministic, slightly stricter). VLMEvalKit's MUIRDataset (image_mcq.py:667) inherits the generic MCQ pipeline with optional GPT extraction, but the authors' own code is LLM-free. Official reporting is overall accuracy plus per-task accuracy (lmms-eval utils.py muir_aggregation)." | |
| }, | |
| "prompt_template": "{{ question }}", | |
| "mapping_from_source": { | |
| "media": { | |
| "from": "images", | |
| "type": "list", | |
| "min_items": 2, | |
| "max_items": 9 | |
| }, | |
| "id": { | |
| "from": "id" | |
| }, | |
| "question": { | |
| "from": "question" | |
| }, | |
| "answer": { | |
| "from": "answer", | |
| "optional": true | |
| }, | |
| "extra": { | |
| "task": { | |
| "from": "task" | |
| } | |
| }, | |
| "source": { | |
| "format": "json", | |
| "url": { | |
| "test": "https://huggingface.co/datasets/MUIRBENCH/MUIRBENCH" | |
| } | |
| } | |
| }, | |
| "prompt_template_source": { | |
| "origin": "official", | |
| "reference": "https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/dataset/image_mcq.py (MUIRDataset.build_prompt — question column already bakes 'Answer with the option letter only.' trailer; bare-question template avoids duplication)", | |
| "notes": "Tier 3: VLMEvalKit MUIRDataset.build_prompt: MuirBench's question column already contains its answer-format trailer baked in, so the official handling is bare {{ question }}." | |
| } | |
| } | |
| } | |
| } |