| { | |
| "name": "R-Bench-V", | |
| "release_date": "2026-05-25", | |
| "subsets": { | |
| "main": { | |
| "language": [ | |
| "en" | |
| ], | |
| "modalities": [ | |
| "single_image_start" | |
| ], | |
| "task_type": "short_answer_qa", | |
| "score_pipeline": [ | |
| "llm-judge" | |
| ], | |
| "score_params": { | |
| "rubric": "binary", | |
| "official_judge_model": "gpt-4o" | |
| }, | |
| "score_protocol": { | |
| "reference": "official paper arXiv:2505.16770 §4.2 — 'we adopt a unified LLM-as-a-Judge framework, with the judge model being GPT-4o. We report Top-1 accuracy (%) as our default evaluation metric.' for all 803 questions (both MCQ and open-ended); official eval code released as a VLMEvalKit fork (github.com/CHEN-Xinsheng/VLMEvalKit_RBench-V). [Quote re-verified verbatim against arxiv.org/html/2505.16770 §4.2, 2026-07-07.]", | |
| "note": "Every sample is officially judge-graded binary; ground truths are often multi-part/unit-laden strings (e.g. '（1）60W（2）80%', '11cm'), so rule pre-resolution is unreliable and was not part of the official protocol. Not present in upstream VLMEvalKit or lmms-eval. NOTE: the math/physics/game/counting splits are category re-slices of the full split (176+157+275+195=803) — pooling any category split with 'full' double-counts; official headline = accuracy over the 803 'full' rows." | |
| }, | |
| "prompt_template": "{% if has_image %}<image>{% endif %}{{ question }}", | |
| "mapping_from_source": { | |
| "media": { | |
| "from": "image", | |
| "type": "list", | |
| "min_items": 0, | |
| "max_items": 1 | |
| }, | |
| "id": { | |
| "from": "id" | |
| }, | |
| "question": { | |
| "from": "question" | |
| }, | |
| "answer": { | |
| "from": "answer", | |
| "optional": true | |
| }, | |
| "extra": { | |
| "catagory": { | |
| "from": "catagory" | |
| }, | |
| "has_image": { | |
| "from": "has_image" | |
| } | |
| }, | |
| "source": { | |
| "format": "huggingface", | |
| "url": { | |
| "full": "https://huggingface.co/datasets/R-Bench/R-Bench-V", | |
| "math": "https://huggingface.co/datasets/R-Bench/R-Bench-V", | |
| "physics": "https://huggingface.co/datasets/R-Bench/R-Bench-V", | |
| "game": "https://huggingface.co/datasets/R-Bench/R-Bench-V", | |
| "counting": "https://huggingface.co/datasets/R-Bench/R-Bench-V" | |
| } | |
| } | |
| }, | |
| "prompt_template_source": { | |
| "origin": "official", | |
| "reference": "https://github.com/SCUT-DLVCLab/R-Bench (R-Bench-V official: bare image+question with has_image conditional)", | |
| "notes": "Tier 1: R-Bench-V official evaluation prompt as published by the dataset authors." | |
| } | |
| } | |
| } | |
| } |