Zero-shot accuracy on ImageNet (in the CLIP setting) is lower than the number reported in the paper

[shyammarjit]

Zero-shot accuracy on ImageNet (in the CLIP setting)

Top-1 accuracy: 77.15
Top-5 accuracy: 95.51

Paper reported accuracy on ImageNet is (Top-1): 88.5

[209ye]

The 88.5 accuracy mentioned in the paper here should be the 10B model, which seems to have not been announced yet. The published 0.4b model looks at the data in the paper and is (Top-1) 78.5.

[shyammarjit]

How do I load the 10B model? Is it open-sourced yet?