DataFlex: Unified Framework for Data-Centric Dynamic LLM Training — 122 HF Upvotes

arXiv / HuggingFace Daily Papers·medium signal

Built on LLaMA-Factory by Peking University and Shanghai AI Lab, DataFlex unifies three paradigms of data-centric training: dynamic sample selection, domain mixture adjustment, and sample reweighting. Provides extensible trainer abstractions as a drop-in replacement for standard LLM training loops. At 122 HuggingFace upvotes, it addresses the growing need for systematic data optimization tooling.