OP-Mix: Unified Data Mixing via Low-Rank Adapter Interpolation
October 30, 2026
OP-Mix treats data mixing as an online decision problem across pretraining, midtraining, and instruction tuning. It simulates candidate mixtures by interpolating between low-rank adapters to optimize composition throughout the training lifecycle.
HOW THIS AFFECTS YOU
●
builderThis offers a path toward more efficient continual learning and adaptive pretraining pipelines.
●
researcherYou can use adapter interpolation to simulate data mixtures without full retraining.