Budget-constrained layer-wise bit protection of INT8 neural network weights against soft errors

This dataset supports the study of budget-constrained selective weight-bit protection in VGG16 and ResNet-18. It includes single-bit fault records, random fault patterns, protection configurations, evaluation results, statistical summaries, and analysis and plotting code. Comparisons with adapted DQN and FEC-TopBits methods are included for VGG16. The files support verification of the reported results and reproduction of the manuscript figures. VGG16 uses INT8 weights and activations, while ResNet-18 uses INT8 weights with FP32 activations and arithmetic.