Optimal Data Mixing for LLM Pretraining Using Mixture Experiments

Yicheng Mao, Hongru Du· August 26, 2026 View original

Key takeaways

  • LLM data mixing can be optimized using classical mixture experiment design principles.
  • Scheffé response-surface models reveal significant interaction effects between data domains.
  • Optimal experimental designs can reduce proxy training runs by approximately 25%.
  • This approach improves statistical efficiency and interpretability of data mixing strategies.

Who benefits

AI/ML EngineeringResearch & DevelopmentCloud ComputingData Science

Summary

This study proposes treating LLM pretraining data mixing as a classical mixture experiment, using response surface methodology and optimal design. It shows that this framework can interpret domain interactions and design more efficient proxy experiments, reducing the need for extensive proxy runs.

Deciding the optimal proportion of data from various domains for large language model pretraining is a critical challenge. Current methods often involve training smaller "proxy" models on different data mixtures, then fitting a response model to predict performance for larger-scale training. This research reframes this process as a classical mixture experiment, where data domains are components and token shares are proportions. By applying sparse second-order Scheffé response-surface models and I-optimal designs, the study demonstrates how to both interpret observed mixture responses and design more statistically efficient proxy experiments. The analysis reveals that the value of data domains is highly relational, with significant pairwise interactions, particularly between specific domains and web-derived text. This approach can reduce the number of necessary proxy runs by about 25% while maintaining accurate mixture rankings across different model scales.

Why it matters

Professionals involved in LLM development can significantly optimize pretraining data strategies, leading to more efficient resource allocation, faster iteration cycles, and potentially better model performance with reduced computational cost.

How to implement this in your domain

  1. 1Adopt a structured experimental design approach for LLM data mixing, treating data domains as mixture components.
  2. 2Utilize response surface methodology, specifically Scheffé models, to analyze the impact of different data proportions.
  3. 3Implement I-optimal designs to select proxy training runs more efficiently, reducing the total number of experiments.
  4. 4Analyze interaction effects between different data domains to uncover non-additive performance gains.
  5. 5Calibrate simulation studies with observed proxy-training responses to validate and refine experimental designs.

Original post by Yicheng Mao, Hongru Du

"arXiv:2608.23922v1 Announce Type: new Abstract: Data mixing is a central design problem in large language model pretraining: given a fixed token budget, practitioners must decide how much data to allocate to each domain. Recent proxy-based methods address this problem by training…"

View on X

Originally posted by Yicheng Mao, Hongru Du on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses