VStarBench / metadata.json
ZTWHHH's picture
metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
548b6b8 verified
Raw
History Blame Contribute Delete
2.75 kB
{
"name": "VStarBench",
"release_date": "2026-05-15",
"subsets": {
"main": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "multiple_choice_qa",
"score_pipeline": [
"rule-match",
"llm-match"
],
"score_protocol": {
"reference": "vlmevalkit@vlmeval/dataset/image_mcq.py:249-324 (ImageMCQDataset.evaluate_heuristic) -> vlmevalkit@vlmeval/dataset/utils/multiple_choice.py:475-508 mcq_vanilla_eval: can_infer rule prefetch, then GPT extractor (extract_answer_from_item @vlmeval/dataset/utils/multiple_choice.py:359-407) maps verbose output to a letter, rule-compare vs GT; VStarBench registered as plain ImageMCQDataset @vlmeval/dataset/image_mcq.py:127. Cross-check: lmms-eval@lmms_eval/tasks/vstar_bench/utils.py:45-108 grades with deterministic regex letter extraction (pure rule), per-category (direct_attributes / relative_position) + overall accuracy",
"note": "Per-category breakdown (direct_attributes/relative_position, in extra.category) is aggregation-level only. Options are baked inline into the question text as '(A) ...' lines with the letter-answer directive; GT is a bare letter (verified via HF rows) — extraction must parse inline parenthesized options."
},
"prompt_template": "<image>{{ question }}",
"mapping_from_source": {
"media": {
"from": "image",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "question_id"
},
"question": {
"from": "text"
},
"answer": {
"from": "label",
"optional": true
},
"extra": {
"category": {
"from": "category"
}
},
"source": {
"format": "json",
"url": {
"test": "https://huggingface.co/datasets/craigwu/vstar_bench"
}
}
},
"prompt_template_source": {
"origin": "source_column",
"reference": "https://huggingface.co/datasets/craigwu/vstar_bench (source `text` column = full pre-rendered prompt: question + (A)/(B)/(C)/(D) options + 'Answer with the option's letter from the given choices directly.' trailer baked in)",
"notes": "Tier 2.2: pass-through {{ question }} where question = upstream `text` column. All 191/191 rows carry pre-rendered prompt with options A-D inline and the answer-letter instruction. Cross-reference: lmms-eval vstar_bench task (https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/lmms_eval/tasks/vstar_bench/utils.py) reformats the same data; mm-eval preserves the source column verbatim."
}
}
}
}