dynamic-data-mixing

RegMix-D extends RegMix to dynamic data mixing by using loss trajectories from proxy runs to predict optimal mixtures at multiple training stages, achieving improvements over static methods.