Data-centric LLM training with dynamic sample selection, domain mixture optimization, and example reweighting inside the LLaMA-Factory training loop.

DataFlex is a data-centric training framework that enhances model performance by either selecting the most influential samples, optimizing their weights, or adjusting their mixing ratios.