Can we use in-context multimodal data for finetuning?

[waltonfuture]

Thanks for your great work! However, it seems that we can only use data that contains one image for SFT. Can we use in-context multimodal data (i.e., containing multiple images) for finetuning?