some confuse for "Pre-trained CLIP-Text encoders for multiple languages"

[moluchase]

if i have <other language text , image, label> pair data, can i directly use 'distilbert-base-multilingual-cased' to pre-train clip model?
Why re-train a model for englist to other languages ?

[ZiboZ]

I guess because the image-text matching task is done on English corpus only.