Details in training the multilingual model and the multimodal model surpporting images and videos?