Unable to reproduce the paper results

[Raion-Shin]

The official weights yield data that is generally lower than that in the paper. In Table 2, for the CLIP SF column, the 16 numbers I measured in order are as follows:
0.4255, 0.7174, 0.1774, 0.8342, 0.5251, 0.777, 0.3871, 0.916, 0.1808, 0.3165, 0.3961, 0.2221, 0.2419, 0.4379, 0.6016, 0.4437
After fine-tuning, the weights I trained myself are only slightly lower than the official weights.