Question about class_text for COCO zero-shot evaluation

[wms405]

Thanks for your great work!
I reproduced the COCO zero-shot results using wedetect-base. When I used coco_zh_class_texts.json as class_text, I obtained 52.1 mAP, which is consistent with the result reported in the paper.
However, when I replaced it with data/texts/coco_class_texts.json without changing any other settings, the performance dropped to 42.3 mAP. Is the 42.3 result reasonable? why switching the class text leads to such a significant performance drop?

[fushh]

Since WeDetect was only trained on Chinese data, its performance degrades when using data/texts/coco_class_texts.json.