Large-scale image-text pre-trained models enable zero-shot classification and provide
consistent accuracy across various data distributions. Nonetheless, optimizing these models …