ConceptCoder数据集
收藏资源简介:
ConceptCoder数据集由爱荷华州立大学团队构建,包含80,204个代码示例,涵盖134种常见弱点枚举(CWE)类型。该数据集聚焦漏洞检测和分支预测两大代码推理任务,其中28,974条专项数据标注了空指针解引用、缓冲区溢出等四类漏洞的7种语义概念。数据通过静态分析工具自动生成概念标签,采用真实场景的不平衡分布设计,旨在提升大语言模型对代码语义的层次化推理能力,在软件安全分析和程序行为预测领域具有重要应用价值。
The ConceptCoder dataset was constructed by a team from Iowa State University, comprising 80,204 code examples covering 134 common Common Weakness Enumeration (CWE) types. This dataset focuses on two code reasoning tasks: vulnerability detection and branch prediction. Among them, 28,974 specialized entries are annotated with 7 semantic concepts across 4 types of vulnerabilities including null pointer dereference and buffer overflow. The concept labels are automatically generated via static analysis tools, and the dataset is designed with an imbalanced distribution that aligns with real-world scenarios. It aims to enhance the hierarchical reasoning ability of Large Language Models (LLMs) regarding code semantics, and holds significant application value in the fields of software security analysis and program behavior prediction.




