anchor
Train a smaller model to match a larger model's output distribution rather than the hard labels.