7个大型鲁棒性基准数据集

本研究构建了7个大型鲁棒性基准数据集，包含96种视觉腐败和87种文本腐败方法，用于评估预训练视觉语言模型在特定领域的适应方法的鲁棒性。数据集涵盖多种视觉和文本的腐败类型，旨在解决模型在实际应用中遇到的分布偏移问题，特别是在自动驾驶系统和临床诊断等安全关键应用中的模型鲁棒性。

This study constructs 7 large-scale robust benchmark datasets, which encompass 96 visual corruption and 87 text corruption methods, for evaluating the robustness of domain-adapted pre-trained vision-language models. The datasets cover diverse visual and text corruption types, aiming to address the distribution shift problem encountered by models in real-world applications, especially the model robustness in safety-critical scenarios such as autonomous driving systems and clinical diagnostics.