sorry, why the sft dataset and sft trainer do not have image input like rl_dataset? The paper show that there should be image for SFT training.