LLaVE is a series of large language and vision embedding models trained on a variety of multimodal embedding datasets