MA-ProofBench
收藏资源简介:
MA-ProofBench是首个用于评估大语言模型在数学分析领域定理证明能力的正式基准数据集。该数据集包含200个经过严格形式化的定理证明问题,使用Lean 4定理证明器和Mathlib数学库(v4.28.0)进行构建。问题根据难度分为两个层级:Level I包含100个本科级别的数学分析基础教材习题;Level II包含100个博士级别的顶尖大学考试问题。数据集涵盖6个核心数学主题和27个子类别,包括实函数、泛函分析、复变函数、测度与积分、算子理论、序列级数与可和性,特别针对先前基准中代表性不足、需要深度推理连续性、极限和拓扑结构的领域。每个数据样本包含唯一ID、难度层级拆分、问题的自然语言描述、带有`sorry`占位符的Lean 4形式化定理陈述、必要的导入头、主题分类、子类别标签以及所依赖的Mathlib版本。数据集通过人主导、LLM辅助的形式化流程构建,并经过独立专家盲审以确保数学保真度,适用于大语言模型在形式化数学和定理证明任务上的评估与微调。
MA-ProofBench is the first formal benchmark dataset designed to evaluate the theorem proving capabilities of large language models in the field of mathematical analysis. It consists of 200 rigorously formalized theorem proving problems, constructed using the Lean 4 theorem prover and the Mathlib mathematical library (v4.28.0). The problems are divided into two difficulty levels: Level I includes 100 undergraduate-level exercises from foundational mathematical analysis textbooks, while Level II contains 100 doctoral-level exam problems from top universities. The dataset covers 6 core mathematical topics and 27 subcategories, including real functions, functional analysis, complex analysis, measure and integration, operator theory, and sequences, series, and summability, with a special focus on areas that are underrepresented in prior benchmarks and require deep reasoning about continuity, limits, and topological structures. Each data sample includes a unique ID, difficulty level split, a natural language description of the problem, a Lean 4 formalized theorem statement with a `sorry` placeholder, necessary import headers, topic classification, subcategory labels, and the dependent Mathlib version. The dataset is built through a human-led, LLM-assisted formalization process and has undergone independent expert blind review to ensure mathematical fidelity, making it suitable for evaluating and fine-tuning large language models on formal mathematics and theorem proving tasks.
数据集概述
MA-ProofBench 是一个用于评估大语言模型(LLM)在数学分析领域定理证明能力的正式基准测试,基于 Lean 4 和 Mathlib (v4.28.0) 形式化构建。
核心特性
- 题目数量:200 道经过严格形式化的定理证明问题。
- 难度分级:分为两个层级。
- Level I(本科水平):100 道基础教科书习题。
- Level II(博士水平):100 道顶尖大学考试问题。
- 覆盖范围:涵盖 6 大核心主题 和 27 个子类别,包括实函数、泛函分析、复变函数、测度与积分、算子理论以及序列与级数。
类别分布(按 MSC 分类)
| 类别 | Level I | Level II |
|---|---|---|
| 实函数 | 44 | 12 |
| 泛函分析 | 15 | 31 |
| 复变函数 | 19 | 16 |
| 测度与积分 | 13 | 17 |
| 算子理论 | 4 | 23 |
| 序列、级数、可和性 | 5 | 1 |
数据字段
| 字段名 | 类型 | 描述 |
|---|---|---|
id |
int | 基准测试中的唯一问题编号 |
split |
string | 层级标识(level1 或 level2) |
informal_statement |
string | 问题的自然语言表述 |
formal_statement |
string | Lean 4 定理形式化陈述(含 sorry 占位符) |
header |
string | 所需的导入/打开语句(通常为 import Mathlib) |
topic |
string | MSC 顶级分类 |
tag |
string | MSC 子类别 |
version |
string | 验证所依据的 Mathlib 版本 |
数据使用示例
python from datasets import load_dataset
ds = load_dataset("openbmb/MA-ProofBench", split="test")
level1 = ds.filter(lambda x: x["split"] == "level1") # 100 个问题 level2 = ds.filter(lambda x: x["split"] == "level2") # 100 个问题
print(ds[0]["formal_statement"])
许可
本项目基于 MIT 许可证 发布。
引用
bibtex @article{ma-proofbench, title={MA-ProofBench: A Two-Tiered Evaluation of LLMs for Theorem Proving in Mathematical Analysis}, author={Lushi Pu and Weiming Zhang and Xinheng Xie and Zixuan Fu and Bingxiang He and Hongya Lyu and Xin Li and Jie Zhou and Yudong Wang}, year={2026}, eprint={2606.13782}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2606.13782}, }




