File size: 3,016 Bytes
7386af3 6f0829c 7386af3 6f0829c | 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 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 | {
"name": "VisIT-Bench",
"release_date": "2023-11-09",
"subsets": {
"default": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "open_ended_qa",
"score_pipeline": [
"llm-judge"
],
"score_params": {
"rubric": "binary",
"official_judge_model": "gpt-4",
"judge_inputs": [
"instruction_conditioned_caption"
]
},
"score_protocol": {
"reference": "Official repo https://github.com/mlfoundations/VisIT-Bench — GPT-4 (reference-free) pairwise judge run in both orderings; leaderboard = Elo ratings + headline win rate vs the GPT-4 reference response (gpt4_prediction). README re-verified 2026-07-07: 'reference-free GPT-4 as an evaluator', 'elo ratings and the win-rate-against reference', two judgements per battle aggregated. Not implemented in VLMEvalKit or lmms-eval.",
"note": "Official headline = win rate vs the GPT-4 reference (plus Elo over multi-model battles), aggregated from both-ordering pairwise judgments — no per-sample absolute rubric exists officially. The mm-eval copy stores the GPT-4 reference as both `answer` and extra.reference_response, enabling per-instance judge-vs-reference grading (rubric=binary: model response wins/ties vs reference), which is an approximation of the win-rate headline; Elo is not reproducible per-sample."
},
"prompt_template": "<image>{{ question }}",
"prompt_template_source": {
"origin": "official",
"reference": "https://huggingface.co/datasets/mlfoundations/VisIT-Bench",
"notes": "Tier 1: VisIT-Bench is judge-scored open-ended instruction-following over real images. Per paper, evaluator is an LLM judge that compares model output to reference_response (gpt4_prediction column); we store the reference as both `answer` (for exact-match-style storage compatibility) and `reference_response` (canonical judge field)."
},
"mapping_from_source": {
"media": {
"from": "images",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "id"
},
"question": {
"from": "question"
},
"answer": {
"from": "answer",
"optional": true
},
"extra": {
"n_images": {
"from": "n_images"
},
"reference_response": {
"from": "reference_response"
},
"instruction_conditioned_caption": {
"from": "instruction_conditioned_caption"
},
"instruction_category": {
"from": "instruction_category"
},
"judge_type": {
"from": "judge_type"
}
},
"source": {
"format": "json",
"url": {
"test": "https://huggingface.co/datasets/mlfoundations/VisIT-Bench"
}
}
}
}
}
} |