Search papers, labs, and topics across Lattice.
This paper introduces DecoupleMix, a systematic framework for optimizing data mixtures for Vision Language Models (VLMs) by decoupling the mixture construction into inter-class and intra-class allocation problems. The approach employs an iterative search for inter-class ratios and a constrained convex optimization for intra-class composition, allowing for principled data selection and validation. Experiments demonstrate that DecoupleMix outperforms heuristic methods and achieves competitive performance with significantly fewer training tokens compared to larger multimodal budgets.
Optimal data ratios discovered through DecoupleMix yield competitive VLM performance with 80B fewer training tokens than traditional methods.
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.
