ZeroBench / metadata.json
ZTWHHH's picture
metadata: 2026-07-07 revised taxonomy (task_type format vocabulary multiple_choice/yes_no/short_answer/free_form; score_type incl. none; re-verified protocol evidence; repaired prompt provenance)
881d6e9 verified
Raw
History Blame Contribute Delete
2.64 kB
{
"name": "ZeroBench",
"release_date": "2025-02-11",
"subsets": {
"default": {
"language": [
"en"
],
"modalities": [
"multi_image_start"
],
"task_type": "short_answer",
"score_type": "rule",
"score_protocol": {
"reference": "official jonathan-roberts1/zerobench@README evaluation snippet — extract final answer from curly braces (regex r\"\\{(.*?)\\}\"), then case-insensitive exact match with length check against ground truth; pass@1 accuracy. lmms-eval@lmms_eval/tasks/zerobench/utils.py:21-56 — normalized exact match with Decimal numeric equality, no LLM. [Re-verified 2026-07-07: README snippet quotes pattern r\"\\{(.*?)\\}\" and parsed_answer[:len(gt)].lower()==gt.strip().lower() with length check, pass@1 and k/n reliability; clone utils.py:22-50 _normalize_text/_to_decimal/_is_exact_match confirmed.]",
"note": "Official prompt instructs 'give the final answer in curly braces'; the mm-eval prompt template is bare images+question with no answer-format instruction (lmms-eval at least appends 'Answer with only the final answer.'), so template extraction has weaker anchors. mm-eval ships both zerobench (100 main) and zerobench_subquestions (334) HF configs under one 'default' subset; official headline is main-question pass@1 (o1-style k/n reliability variants and subquestion accuracy reported separately) — do not pool the two configs into one accuracy number."
},
"prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}",
"prompt_template_source": {
"origin": "official",
"reference": "https://huggingface.co/datasets/jonathan-roberts1/zerobench",
"notes": "Tier 1: ZeroBench published as open-ended VQA on HF; dataset card describes single-prompt evaluation across both zerobench (100 main) and zerobench_subquestions (334 decomposed) splits."
},
"mapping_from_source": {
"media": {
"from": "question_images_decoded",
"type": "list",
"min_items": 1,
"max_items": 3
},
"id": {
"from": "question_id"
},
"question": {
"from": "question_text"
},
"answer": {
"from": "question_answer",
"optional": true
},
"source": {
"format": "huggingface",
"url": {
"zerobench": "https://huggingface.co/datasets/jonathan-roberts1/zerobench",
"zerobench_subquestions": "https://huggingface.co/datasets/jonathan-roberts1/zerobench"
}
}
}
}
}
}