Rethinking Data Mixing from the Perspective of Large Language Models
This work addresses theoretical gaps in data mixing for LLM training, which is crucial for improving model generalization, though it appears incremental as it builds on existing empirical methods.
The paper tackles fundamental questions about data mixing strategies for large language model training by establishing theoretical connections between gradient dynamics and domain distributions, and introduces DoGraph, a reweighting framework that formulates data scheduling as a graph-constrained optimization problem. Experiments on GPT-2 models show that DoGraph consistently achieves competitive performance.
Data mixing strategy is essential for large language model (LLM) training. Empirical evidence shows that inappropriate strategies can significantly reduce generalization. Although recent methods have improved empirical performance, several fundamental questions remain open: what constitutes a domain, whether human and model perceptions of domains are aligned, and how domain weighting influences generalization. We address these questions by establishing formal connections between gradient dynamics and domain distributions, offering a theoretical framework that clarifies the role of domains in training dynamics. Building on this analysis, we introduce DoGraph, a reweighting framework that formulates data scheduling as a graph-constrained optimization problem. Extensive experiments on GPT-2 models of varying scales demonstrate that DoGraph consistently achieves competitive performance.