Training that mixes a small labeled set with a large unlabeled set, leveraging pseudo‑labels or consistency losses for better generalization.