SCITUNE: Aligning Large Language Models with Human-Curated Scientific Multimodal Instructions