Distillation imo

Big models are accurate but slow and expensive to deploy. Knowledge distillation transfers their capability into a compact model by having the student learn from the teacher's outputs rather than only from hard labels. The key insight, from Hinton and colleagues, is that a teacher's full probability distribution carries 'dark knowledge': even when it predicts 'dog', the relative probabilities for 'wolf' versus 'car' reveal how the teacher sees similarities. Softening these probabilities with a temperature exposes that structure, and the student is trained to match it, often alongside the true labels. The goal is an efficient student that benefits from teacher supervision; the actual accuracy and speed must be measured for the task and implementation. DistilBERT and TinyBERT are well-known distilled language models.