26PaperSubmission/repo-vul

RepoVul是一个用于仓库级漏洞检测的综合数据集，专门用于微调轻量级大型语言模型（LLMs）在漏洞检测任务上。该子集包含训练、验证和测试数据，支持Java和Python两种编程语言。任务为二元分类（有漏洞代码 vs 无漏洞代码），数据格式采用指令调优提示格式，标签为0（无漏洞）或1（有漏洞）。每个示例包括漏洞标签、完整提示（包含指令、代码片段和仓库上下文，如文件位置、调用者和被调用者函数）、用于映射到CVE和仓库的标识符以及函数名。数据集覆盖多种CWE类别，如跨站脚本（XSS）、SQL注入、缓冲区错误等，适用于微调、漏洞检测、代码分析和基准测试。RepoVul is a comprehensive dataset for repository-level vulnerability detection. This subset contains training, validation, and test data for fine-tuning lightweight Large Language Models (LLMs) on vulnerability detection tasks. It supports Java and Python languages, with a binary classification task (vulnerable vs non-vulnerable code) in instruction-tuning prompt format. Labels are 0 (non-vulnerable) or 1 (vulnerable). Each example includes a vulnerability label, a complete prompt with instruction, code snippet, and repository context (such as file location, caller and callee functions), an identifier for mapping to CVE and repository, and the function name. The dataset covers various CWE categories including Cross-site Scripting (XSS), SQL Injection, buffer errors, etc., and is suitable for fine-tuning, vulnerability detection, code analysis, and benchmarking.