SFT_Dataset
收藏资源简介:
Pythagoras SFT 数据集是一个用于定理证明与形式推理的大规模合成数据集,专为基于Lean4的形式化证明任务而设计。该数据集旨在通过提供自然语言问题、对应的Lean4形式化陈述、形式化证明以及推理链,来支持高效的形式证明模型训练。数据集包含约841,000个问题-证明对,当前公开发布了126,000个实例的子集,按难度分为简单(30,000个)、中等(49,000个)和困难(47,000个)三个级别。每个数据实例包含以下字段:自然语言问题描述(Question)、问题来源家族(Source)、Lean4形式化陈述(Formal Statement)、Lean4形式化证明(Formal Proof)、问题所属的广泛数学类别(Problem Type)、证明前的推理链或证明计划(Reasoning Chain)以及难度标签(Difficulty)。该数据集适用于文本生成任务,特别是自动化定理证明、形式推理以及Lean4代码生成等场景。数据生成流程的完整说明可在关联论文《Pythagoras-Prover: Advancing Efficient Formal Proving via Augmented Lean Formalisation》中找到。
The Pythagoras SFT dataset is a large-scale synthetic dataset for theorem proving and formal reasoning, specifically designed for formal proof tasks based on Lean4. It aims to support efficient formal proof model training by providing natural language problems, corresponding Lean4 formal statements, formal proofs, and reasoning chains. The dataset contains approximately 841,000 problem-proof pairs, with a publicly released subset of 126,000 instances categorized into three difficulty levels: easy (30,000), medium (49,000), and hard (47,000). Each data instance includes the following fields: natural language problem description (Question), problem source family (Source), Lean4 formal statement (Formal Statement), Lean4 formal proof (Formal Proof), broad mathematical category of the problem (Problem Type), reasoning chain or proof plan before proof (Reasoning Chain), and difficulty label (Difficulty). This dataset is suitable for text generation tasks, particularly automated theorem proving, formal reasoning, and Lean4 code generation scenarios. A complete description of the data generation process can be found in the associated paper Pythagoras-Prover: Advancing Efficient Formal Proving via Augmented Lean Formalisation.
数据集概述
基本信息
- 数据集名称:Pythagoras-Prover SFT Dataset
- 语言:英语(en)
- 许可证:Apache-2.0
- 规模:100K < n < 1M
- 任务类别:文本生成(text-generation)
- 标签:Lean4、定理证明、形式推理
数据集内容
该数据集包含约 841K 个问题,每个问题配有:
- Lean 形式化陈述(Formal Statement)
- 形式化证明(Formal Proof)
- 推理链(Reasoning Chain)
当前发布的子集包含 126K 个实例,按难度分为:
- 30K 简单实例(easy)
- 49K 中等实例(medium)
- 47K 困难实例(hard)
完整数据集将后续发布。
数据格式
每一行数据包含以下字段:
- Question:自然语言问题描述
- Source:问题的来源类别
- Formal Statement:Lean 4 形式化表述
- Formal Proof:针对形式化陈述生成的 Lean 4 证明
- Problem Type:数学问题的广义类别
- Reasoning Chain:形式化证明之前的证明计划或推理过程
- Difficulty:难度标签(easy / medium / hard)
示例
一个典型数据条目包含一个组合学问题、其 Lean 4 形式化陈述与证明、推理链,以及难度标签为“easy”。
引用
bibtex @article{leang2026pythagoras, title={Pythagoras-Prover: Advancing Efficient Formal Proving via Augmented Lean Formalisation}, author={Leang, Joshua Ong Jun and Zhao, Zheng and Stoian, Mihaela Catalina and Xu, Qiyuan and Li, Haonan and Li, Wenda and Cohen, Shay B. and Giunchiglia, Eleonora}, journal={arXiv preprint arXiv:2606.12594}, year={2026} }
相关链接
- 项目主页:https://pythagoras-lm.github.io/
- GitHub 仓库:https://github.com/Pythagoras-LM/Pythagoras-Prover
- 论文:https://huggingface.co/papers/2606.12594




