MindLLM: A Subject-Agnostic and Versatile Model for fMRI-to-Text Decoding