cannot achieve the claimed zero-shot retrieval results of SigLIP2 ViT/B16-224 on COCO and Flickr30K in your paper

On COCO and Flickr30K, the zero-shot SigLIP2 ViT/B16-224 model fails to achieve the retrieval performance claimed in the paper: in the paper, COCO (t2i-R@1: 52.1, i2t-R@1: 68.9) and Flickr30K (t2i-R@1: 80.7, i2t-R@1: 93.0). However, in my own experiments, the results are COCO (t2i-R@1: 48.58, i2t-R@1: 64.96) and Flickr30K (t2i-R@1: 73.66, i2t-R@1: 89.30).