Higher Accuracy Observed with clip_benchmark

[ReubenDo]

Hello,

First, congratulations on your excellent work!

I used the ViT-H-14-CLIPA-336-laion2B model from Hugging Face with CLIP Benchmark and obtained higher scores on imagenet1k than those reported in your arXiv paper.

Here are the results I obtained:

Metric	Precision	Recall	F1-Score	Support
Accuracy			0.864	10000
Macro Avg	0.871	0.864	0.862	10000
Weighted Avg	0.871	0.864	0.862	10000

These scores are higher than those reported in the paper.

Do you know what might explain these differences?

Best regards,
Reuben