| { |
| "name": "OlympiadBench", |
| "release_date": "2024-06-17", |
| "subsets": { |
| "main": { |
| "language": [ |
| "en", |
| "zh" |
| ], |
| "modalities": [ |
| "multi_image_start", |
| "text" |
| ], |
| "task_type": "short_answer_qa", |
| "score_pipeline": [ |
| "exact-match", |
| "rule-match" |
| ], |
| "score_protocol": { |
| "reference": "official OpenBMB/OlympiadBench@eval/auto_scoring_judge.py (MathJudger) — deterministic sympy/latex math-expression equivalence with per-item precision; vendored at lmms-eval@lmms_eval/tasks/olympiadbench/olympiadbench_evals.py:21-72 (OlympiadBenchEvaluator.judge, default precision 1e-8) used in en_utils.py:51-66, and at VLMEvalKit@vlmeval/dataset/utils/olympiadbench.py:252-471 (MathJudger, default non-API path). [Re-verified in clone 2026-07-07: default self.precision=1e-8 at olympiadbench_evals.py:39, judge at :72; en_utils.py:52 precision=doc['error']; MathJudger class at olympiadbench.py:252, judge at :300.]", |
| "note": "Official judge takes a per-item numeric precision from the dataset's `error` field (lmms-eval en_utils.py:52 precision=doc['error']); the mm-eval copy does not ship that field, so the 1e-8 default (effectively exact) applies — numeric answers with stated tolerance may be under-scored. mm-eval ships the FULL 8476-row release across 18 configs (OE/TP x multimodal(MM)/text-only(TO) x maths/physics x en/zh; TO splits are text-only, modalities tag 'text'). The 1748 TP theorem-proving rows ship EMPTY answers and are not auto-gradable — the official protocol only collects submissions for them (lmms-eval en_utils.py:57-59); per-sample rule scoring applies to the 6728 OE rows only. Prompt instructs boxed final answer, enabling template extraction. Scorer needs math-expression equivalence for LaTeX answers (e.g. '$\\frac{1}{2}$')." |
| }, |
| "prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{% if language == 'English' %}The following is an open-ended problem from an International {% if subject == 'Physics' %}Physics{% else %}Mathematics{% endif %} competition. Please calculate the answer according to the given requirements and the information provided. Please use LaTeX format to represent the variables and formulas used in the solution process and results. Please end your solution with \"So the final answer is \\boxed{answer}.\" and give the result explicitly.{% else %}以下是中国{% if subject == 'Physics' %}物理{% else %}数学{% endif %}竞赛中的解答题。请根据题目的要求和所提供的信息计算得出答案。解答过程和结果中使用的变量和公式请使用LaTeX格式表示。请在最后以\"所以最终答案是\\boxed{答案}。\"显式给出结果。{% endif %}\n{{ question }}{% if context %}\n\nContext: {{ context }}{% endif %}", |
| "mapping_from_source": { |
| "media": { |
| "from": "images", |
| "type": "list", |
| "min_items": 0, |
| "max_items": 4 |
| }, |
| "id": { |
| "from": "id" |
| }, |
| "question": { |
| "from": "question" |
| }, |
| "answer": { |
| "from": "answer", |
| "optional": true |
| }, |
| "extra": { |
| "n_images": { |
| "from": "n_images" |
| }, |
| "context": { |
| "from": "context" |
| }, |
| "subject": { |
| "from": "subject" |
| }, |
| "subfield": { |
| "from": "subfield" |
| }, |
| "difficulty": { |
| "from": "difficulty" |
| }, |
| "language": { |
| "from": "language" |
| }, |
| "answer_type": { |
| "from": "answer_type" |
| } |
| }, |
| "source": { |
| "format": "json", |
| "url": { |
| "OE_MM_maths_en_COMP": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "OE_MM_maths_zh_CEE": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "OE_MM_maths_zh_COMP": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "OE_MM_physics_en_COMP": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "OE_MM_physics_zh_CEE": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "OE_TO_maths_en_COMP": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "OE_TO_maths_zh_CEE": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "OE_TO_maths_zh_COMP": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "OE_TO_physics_en_COMP": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "OE_TO_physics_zh_CEE": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "TP_MM_maths_en_COMP": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "TP_MM_maths_zh_CEE": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "TP_MM_maths_zh_COMP": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "TP_MM_physics_en_COMP": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "TP_TO_maths_en_COMP": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "TP_TO_maths_zh_CEE": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "TP_TO_maths_zh_COMP": "https://huggingface.co/datasets/Hothan/OlympiadBench", |
| "TP_TO_physics_en_COMP": "https://huggingface.co/datasets/Hothan/OlympiadBench" |
| } |
| } |
| }, |
| "prompt_template_source": { |
| "origin": "official", |
| "reference": "https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/dataset/image_vqa.py (OlympiadBench.build_prompt — language×subject×is_theorem_proving switching, English/Chinese preambles)", |
| "notes": "Tier 3: VLMEvalKit OlympiadBench.build_prompt — per-row language/subject branching for international Olympiad-style problems." |
| } |
| } |
| } |
| } |
