VulnSage

该数据集是一个由多种大型开源C/C++系统软件项目精心策划的全面评估框架和数据库。它采用启发式噪声预过滤方法，并结合了基于LLM的推理，以确保漏洞的代表性并最小化噪声。此外，该框架支持在函数、文件以及跨函数级别进行多粒度分析，并采用了四种不同的零样本提示策略：基础、思维链、思考以及思考与验证。该任务旨在进行软件漏洞检测和补丁验证。

This dataset is a comprehensively curated evaluation framework and database consisting of multiple large-scale open-source C/C++ system software projects. It adopts heuristic noise pre-filtering methods combined with LLM-based reasoning to ensure the representativeness of the included vulnerabilities and minimize irrelevant noise. Furthermore, the framework supports multi-granularity analysis at the function, file, and cross-function levels, and employs four distinct zero-shot prompting strategies: Basic, Chain-of-Thought, Reflection, and Reflection with Verification. This task is designed for software vulnerability detection and patch verification.