I have the feeling alot of models include alot of data in many languages. Would it make more sense to train just on english data and have a seperate translation layer? Or do i misunderstand something?