Zubnet AILearnWiki › Distillation
Training

Distillation

Knowledge Distillation, Model Distillation
Training a smaller "student" model to mimic the behavior of a larger "teacher" model. Instead of training the student on raw data with hard labels (cat/dog), you train it on the teacher's soft probability distributions (70% cat, 20% dog, 10% fox). The soft outputs carry more information than hard labels because they encode the teacher's uncertainty and the relationships between categories.

Why it matters

Distillation is how the industry makes powerful AI accessible. A 70B parameter model might be too large and expensive for real-time applications, but a 7B model distilled from it can capture 90% of the capability at 10% of the cost. Many of the small, fast models people run locally are distilled from larger frontier models.

Deep Dive

The original insight from Hinton et al. (2015) was that a teacher's output probabilities contain "dark knowledge" — information about which wrong answers are almost right. A digit classifier that sees a "7" might output 0.8 for "7" but 0.15 for "1" and 0.03 for "9" — revealing that 7s look more like 1s than 9s. A student trained on these soft targets learns these relationships, which hard labels ("it's a 7, period") don't convey.

In the LLM Era

For LLMs, distillation takes several forms. The most common is training a smaller model on outputs generated by a larger model — you run the teacher on a large set of prompts, collect its responses, and fine-tune the student on those (prompt, response) pairs. This is sometimes called "distillation through generation." It's controversial because some model licenses prohibit using outputs to train competing models, and because it can create models that sound confident but lack the teacher's deeper reasoning abilities.

Distillation vs. Quantization

People sometimes confuse distillation with quantization. Quantization shrinks a model by reducing numerical precision (32-bit to 4-bit) — same model, smaller numbers. Distillation creates an entirely new, architecturally smaller model — fewer layers, smaller dimensions — that has learned from the teacher. They're complementary: you can distill a 70B model into a 7B model and then quantize the 7B model to make it even smaller.

Related Concepts

← All Terms
← Diffusion Transformer Distributed Training →