Could you kindly share the scripts you used to evaluate the zero-shot accuracy of mlcd-ViT-bigG-14? We are currently preparing an analysis paper that examines the gap between image classification accuracy and post-LLaVA fine-tuning performance. To support this, we are particularly interested in the zero-shot performance results of mlcd-ViT-bigG-14.