Data-centric LLM training framework inside the LLaMA-Factory loop, featuring dynamic sample selection, domain mixture optimization and adaptive efficient example reweighting schedules.