Why 10 classification results for CLIP is lower than 10%?

[blacksheepwatch]

In Table 2. Experimental results on DomainNet dataset with feature & Label Shift, the results for Zero-Shot CLIP, PromptFL, PromptFL+FedProx are so strange and even totally wrong.
In your GitHub repository and your work in both this one and ICML 2024: Harmonizing Generalization and Personalization in Federated Prompt Learning, you mention that you use the 10 classes for the task. But as you present, the accuracy is 8.72±1.73 on the Clipart for the Zero-Shot CLIP.

You are really a code genius.