ReForm-Python2Dafny-Dataset
收藏资源简介:
该数据集包含与使用强化学习在大型语言模型中进行形式软件验证的论文相关的数据集。主要包括两个数据集:Python2Dafny和DafnyComp。Python2Dafny数据集包含将Python程序映射到Dafny代码的18000个样本,主要用于监督微调阶段。DafnyComp数据集是一个包含300个具有自动形式化规范的组合形式程序的基准,用于规范推理和评估泛化能力。这两个数据集都提供了JSON格式。
This dataset contains resources associated with the paper on formal software verification using reinforcement learning with large language models. It primarily includes two datasets: Python2Dafny and DafnyComp. The Python2Dafny dataset contains 18,000 samples that map Python programs to Dafny code, and is mainly used for the supervised fine-tuning stage. The DafnyComp dataset is a benchmark comprising 300 compositional formal programs with automated formal specifications, which is employed for specification reasoning and generalization capability evaluation. Both datasets are provided in JSON format.
Re:Form Datasets 数据集概述
数据集基本信息
- 任务类别: 文本生成
- 标签: 代码生成、形式化验证、Dafny、强化学习
数据集来源
- 关联论文: "Re:Form -- Reducing Human Priors in Scalable Formal Software Verification with RL in LLMs: A Preliminary Study on Dafny"
- 项目主页: https://veri-code.github.io/ReForm-page
- 代码仓库: https://github.com/Veri-Code/Veri-Code
数据集内容
包含两个主要数据集:
-
Python2Dafny
- 规模: 18,000个样本
- 用途: 用于监督微调(SFT)阶段
- 内容: Python程序到Dafny代码的映射
-
DafnyComp
- 规模: 300个组合式形式化程序
- 特点: 包含自动形式化的规范
- 用途: 规范推理和评估泛化能力
数据格式
- 所有数据集均以JSON格式提供
使用说明
- 预处理脚本:
python -m src.data_preprocess - 详细使用指南请参考代码仓库文档
引用信息
bibtex @misc{yan2025reformreducinghuman, title={Re:Form -- Reducing Human Priors in Scalable Formal Software Verification with RL in LLMs: A Preliminary Study on Dafny}, author={Chuanhao Yan and Fengdi Che and Xuhan Huang and Xu Xu and Xin Li and Yizhi Li and Xingwei Qu and Jingzhe Shi and Zhuangzhuang He and Chenghua Lin and Yaodong Yang and Binhang Yuan and Hang Zhao and Yu Qiao and Bowen Zhou and Jie Fu}, year={2025}, eprint={2507.16331}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2507.16331}, }




