NeuroCodeBench

NeuroCodeBench是由曼彻斯特大学创建的一个用于软件验证的神经网络基准数据集。该数据集包含32个用纯C语言编写的神经网络，涵盖了数学库、激活函数、错误校正网络、转移函数近似、概率密度估计和强化学习等6个类别，共计607个安全属性。数据集的创建过程涉及将高层次神经网络规范转换为独立的C代码，并确保安全属性的判决已知且平衡。NeuroCodeBench主要用于评估现有软件验证工具的性能，特别是在处理大型神经网络和标准C数学库时的能力，旨在解决安全关键系统中神经网络实现的软件故障问题。

NeuroCodeBench is a neural network benchmark dataset for software verification, created by the University of Manchester. This dataset comprises 32 neural networks written in pure C language, covering six categories: mathematical libraries, activation functions, error correction networks, transfer function approximations, probability density estimation, and reinforcement learning, with a total of 607 safety properties. The dataset creation process involves converting high-level neural network specifications into standalone C code, while ensuring that the verdicts for all safety properties are both known and balanced. NeuroCodeBench is primarily used to evaluate the performance of existing software verification tools, particularly their capabilities when handling large neural networks and standard C mathematical libraries, with the goal of addressing software fault issues related to neural network implementations in safety-critical systems.