ejschwartz/idioms-realtype

REALTYPE数据集是一个为神经反编译研究构建的综合数据集，包含了从GitHub上的真实世界C代码库中提取的154,301个训练函数、540个验证函数和2,322个测试函数。该数据集特别强调包含了大量用户定义的类型（UDTs），并为每个函数提供了反编译代码、原始源代码、用户定义类型的完整定义和调用图信息。数据集分为按函数组织的by_function视图和按二进制组织的by_binary视图，后者不包含调用图或未匹配的函数。

The REALTYPE dataset is a comprehensive dataset constructed for neural decompilation research, containing 154,301 training functions, 540 validation functions, and 2,322 test functions extracted from real-world C code repositories on GitHub. The dataset emphasizes the inclusion of a large number of user-defined types (UDTs) and provides for each function decompiled code, original source code, complete definitions of user-defined types, and call graph information. The dataset is organized into a by_function view that groups data by function and a by_binary view that groups data by binary, with the latter omitting call graphs and unmatched functions.