A preprocessing technique that removes or corrects low-quality or mismatched training examples before training, improving model reliability.