Inconsistency of YOLO-World-S's zero-shot performance on LVIS in the paper

[kaijinz0228]

There seems to be a difference between the zero-shot performance of yolo-world-s on LVIS in Table 2 and Table 9 in the paper. The training data and model scale seems to be the same. Could you clarify any training differences that result in this gap?