Training a model to predict outputs using ground-truth inputs from a teacher model rather than its own predictions.