In machine learning, efficiently leveraging large datasets is essential, particularly when computational resources are limited. Sample compression techniques, which involve using carefully chosen subsets of data to train models, provide significant reductions in runtime and memory requirements while preserving generalization performance. This paper investigates a concept known as “super-teaching,” where a knowledgeable teacher selectively provides an optimal subset of training data to a learner, thereby enhancing learning outcomes. Building on the work of Ma et al. (2018), who demonstrated that a teacher with complete knowledge of the data distribution could significantly improve a learner’s performance, our study extends this idea beyond the scenario in their paper. We present a more robust and generalized approach, offering insights into how super-teaching can be effectively applied to a broader range of machine learning problems, potentially leading to better training efficiency and improved generalization.