livek12bench
收藏资源简介:
LiveK12Bench是一个双语(中文/英文)的K12高中教育基准数据集,专门用于评估模型在STEM(科学、技术、工程和数学)领域的学科能力。该数据集覆盖数学、物理、化学和生物四个核心科目,所有问题均源自真实的中国高中考试试卷,并提供对应的中文和英文版本。数据集包含四个划分:zh_2603、zh_2605、en_2603和en_2605,每个数据样本包括唯一标识符、批次、科目、问题类型、分值、问题文本(含LaTeX内联数学公式)、答案列表、分步解决方案、知识点、图像列表和子集标签。总共有2,114个问题样本,其中1,096个为纯文本问题,1,018个为图文结合问题,2,114个以“仅图像”模式呈现;问题类型包括1,473个选择题、164个填空题和477个问答题;科目分布为数学772个(36.5%)、物理396个(18.7%)、化学532个(25.2%)、生物414个(19.6%)。该数据集适用于教育问答基准测试、多模态理解评估、多步推理能力评估和跨语言模型对比研究,采用CC BY-NC 4.0许可,仅限非商业的研究和教育用途。
LiveK12Bench is a bilingual (Chinese/English) K12 high school education benchmark dataset specifically designed to evaluate model capabilities in STEM (Science, Technology, Engineering, and Mathematics) subjects. It covers four core subjects: mathematics, physics, chemistry, and biology, with all questions sourced from real Chinese high school exam papers and provided in both Chinese and English versions. The dataset includes four splits: zh_2603, zh_2605, en_2603, and en_2605. Each data sample contains multiple structured fields: unique identifier, batch, subject, question type (e.g., multiple-choice, fill-in-the-blank, essay/proof), point value, question text (including LaTeX inline math formulas), answer list, step-by-step solution, knowledge points, image list, and subset labels for special attributes. In total, there are 2,114 question samples, comprising 1,096 text-only questions, 1,018 text-image combined questions, and 2,114 presented in image-only mode from original exams. The question type distribution includes 1,473 multiple-choice questions, 164 fill-in-the-blank questions, and 477 essay questions. Subject distribution is mathematics 772 (36.5%), physics 396 (18.7%), chemistry 532 (25.2%), and biology 414 (19.6%). This dataset is suitable for educational question-answering benchmarks, multimodal understanding (text and image) evaluation, multi-step reasoning assessment, and cross-language (Chinese-English) model comparison studies. It is licensed under CC BY-NC 4.0 for non-commercial research and educational use only.
数据集名称
LiveK12Bench
数据集概述
LiveK12Bench 是一个双语(中文/英文)K12 高中基准数据集,涵盖四个 STEM 科目:数学、物理、化学和生物学。所有题目均来自真实的中国高考或模拟考试试卷,并提供了中英文双语版本。
语言与许可证
- 语言:中文、英文
- 许可证:CC BY-NC 4.0(仅限非商业研究与教育用途)
数据集规模
- 总样本数:2,114 个问题
- 数据分片:共 4 个分片,总计约 4228 条记录(每个分片包含对应的题目)
- 数据集总大小:约 310 MB(下载大小约 315 MB)
数据分片
| 分片名称 | 语言 | 描述 | 样本数 |
|---|---|---|---|
zh_2603 |
中文 | 2026年3月批次 | 1,487 |
zh_2605 |
中文 | 2026年5月批次 | 627 |
en_2603 |
英文 | 2026年3月批次 | 1,487 |
en_2605 |
英文 | 2026年5月批次 | 627 |
数据集特征
每条记录包含以下字段:
id:字符串,唯一标识符(如zh_2603_math_0001)set:字符串,来源批次("2603"或"2605")subject:字符串,科目(biology/chemistry/math/physics)question_type:字符串,题型(选择题、填空题、解答题、证明题 / Multiple Choice、Fill in the Blank、Open-ended、Proof)point_value:整数,题目分值question:字符串,题目文本(LaTeX 内联数学公式用$...$表示)answer:字符串列表,答案(选择题如["A"],填空题/解答题为文本)solution:字符串,详细解题步骤knowledge_points:字符串,涉及的知识点或学习目标images:图像列表,引用的图表/图片(若无则为空列表)subset:字符串列表,子集标签(如complex_layout、rigorous_process、long_reason,仅set=2603的题目有标签)
统计信息
| 类别 | 总计 | 数学 | 物理 | 化学 | 生物学 |
|---|---|---|---|---|---|
| 任务模态 | |||||
| 纯文本 (TO) | 1,096 | 617 (56.3%) | 65 (5.9%) | 240 (21.9%) | 174 (15.9%) |
| 文本+图像 (TI) | 1,018 | 155 (15.2%) | 331 (32.5%) | 292 (28.7%) | 240 (23.6%) |
| 仅图像 (IO, 考试模式) | 2,114 | 772 (36.5%) | 220 (10.4%) | 532 (25.2%) | 414 (19.6%) |
| 题型 | |||||
| 选择题 (MCQ) | 1,473 | 444 (30.1%) | 274 (18.6%) | 419 (28.4%) | 336 (22.8%) |
| 填空题 (FIB) | 164 | 119 (72.6%) | 26 (15.9%) | 18 (11.0%) | 1 (0.6%) |
| 问答题 (Q&A) | 477 | 209 (43.8%) | 96 (20.1%) | 95 (19.9%) | 77 (16.1%) |
| 总计 | 2,114 | 772 (36.5%) | 396 (18.7%) | 532 (25.2%) | 414 (19.6%) |
使用方式
可通过 Hugging Face datasets 库加载:
python
from datasets import load_dataset
ds = load_dataset("Shawn-wxh/livek12bench")
引用
bibtex @misc{livek12bench2026, title = {LiveK12Bench}, author = {Wang, Xiaohan and Yin, Mingze and Zhao, Yilin and Sinbadliu and Li, Dian}, year = {2026}, url = {https://github.com/QQ-MM/LiveK12Bench} } @misc{wang2026livek12bench, title={LiveK12Bench: Have Large Multimodal Models Truly Conquered High School-level Examinations?}, author={Xiaohan Wang and Mingze Yin and Yilin Zhao and Gang Liu and Dian Li}, year={2026}, archivePrefix={arXiv}, url={https://arxiv.org/abs/2605.26781}, }





