The paper proposes a data-centric paradigm for developing an ideal training dataset using the DR4SR framework. It demonstrates significant performance improvements across four datasets.