Sample selection and ordering matter a lot in LLM training: DataFlex makes data scheduling dynamic

Puzzleheaded_Box2842 · reddit · 2026-09-11

The author argues that as training data scales, simple shuffling is no longer enough: noisy/redundant samples, domain imbalance, and example ordering all shape what LLMs learn. They outline four dynamic scheduling levers—dynamic selection (using loss, gradient similarity, or offline scores to pick samples per training window), dynamic reordering for curriculum-style training, dynamic mixing of domain proportions, and dynamic weighting of gradient contributions. The core idea: treat data scheduling as part of optimization rather than a fixed pre-training decision. This is implemented in the open-source OpenDCAI/DataFlex project.

Original post →

More from Research

Research channel →