| title | DecoupleMix: Decoupled Ratio Search and Convex Allocation for Scalable VLM Data Recipes |
|---|---|
| authors | Jiahao Xie, Zhongbin Guo, Qianle Wang, Ruiqi Lu, Dongling Xiao, Wanxuan Sun, Cheng Yang |
| date | 2026-07-27 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2607.24516 |
| upvotes | 7 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf |
| abstract | While data curation for Vision Language Models (VLMs) is increasingly active, public practice for constructing pretraining mixtures remains largely heuristic: practitioners stack datasets that pass quality filters, set cross-domain ratios by intuition, and lack a principled, attributable criterion for admitting new data, while frontier recipes remain undisclosed. We formulate data construction as a systematic mixture-optimization problem and turn it into a reproducible engineering discipline by decoupling the mixture into two orthogonal sub-problems: inter-class ratios across capabilities and intra-class ratios within a category. For inter-class allocation, we use a single-variable iterative search; for intra-class composition, we apply a multidimensional, dataset-level assessment scoring Quality and Difficulty, and formulate selection as a constrained convex optimization with a diversity objective. The DecoupleMix framework delivers two critical capabilities: guiding what data to collect next and rendering dataset validation a controlled, attributable experiment. Experiments show our approach consistently surpasses heuristic baselines. Moreover, optimal ratios discovered on small-scale proxies transfer seamlessly to larger scales without retuning. Using 80B additional multimodal continue-pretraining tokens, our VLM is competitive with strong open-source models trained with substantially larger multimodal budgets. |
cite
@misc{xie2026decouplemix,
title={DecoupleMix: Decoupled Ratio Search and Convex Allocation for Scalable VLM Data Recipes},
author={Jiahao Xie and Zhongbin Guo and Qianle Wang and Ruiqi Lu and Dongling Xiao and Wanxuan Sun and Cheng Yang},
year={2026},
eprint={2607.24516},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2607.24516}
}