Use model output as training data. For better performance you can get some top log probs and minimize kl divergence.