Okay, they simply used an instruction-based training method derived from LLaVA. It's classic and common to leverage previous research methods to keep advancing the field.

→ View original post on X — @dfintelligence