SoundnessBench/SoundnessBench

SoundnessBench数据集包含26个模型，涵盖9种不同的神经网络架构，具有不同的输入尺寸和扰动半径。数据集包含每个实例的VNN-LIB格式形式规范、超时值、模型权重、模型名称和输入形状。此外，还提供了如何将数据集转换为用于神经网络验证器基准测试的格式的说明。The SoundnessBench dataset contains 26 models across 9 distinct neural network architectures with various input sizes and perturbation radii. Each instance includes a VNN-LIB format formal specification, timeout value, model weights, model name, and input shape. Additionally, instructions for converting the dataset into a format suitable for benchmarking neural network verifiers are provided.