SecRepoBench
收藏资源简介:
SecRepoBench 是一个用于评估大型语言模型在真实世界代码库中生成安全代码的基准数据集。该数据集包含了来自27个C/C++代码库的318个代码生成任务,涵盖了15种常见弱点枚举(CWEs)。SecRepoBench 通过利用已知的安全漏洞构建任务,旨在评估LLMs生成安全代码的能力。它采用了动态测试方法来评估LLMs生成代码的正确性和安全性,为研究者提供了研究LLMs在真实世界代码生成中的能力提供了重要的数据资源。
SecRepoBench is a benchmark dataset for evaluating large language models' (LLMs') ability to generate secure code within real-world codebases. This dataset contains 318 code generation tasks sourced from 27 C/C++ code repositories, covering 15 common Common Weakness Enumerations (CWEs). Constructed using known security vulnerabilities, SecRepoBench is designed to assess the capability of LLMs to generate secure code. It adopts dynamic testing methods to evaluate the correctness and security of code generated by LLMs, providing researchers with a critical data resource for studying LLMs' performance in real-world code generation scenarios.




