What can we do to help these models improve? The translations are pretty rough right now