Multi-modal SFT data selection method that first scales to million-level datapool, achieving 99.1% perf with 30% of LLaVA-OVSI. (in construction)