Towards Harmless Multimodal Assistants with Blind Preference Optimization