Distillation

• Key points are in the visual diagram above.

Deploy smaller models for latency/cost while preserving quality. Common in production model routing.

Distillation is a master chef teaching an apprentice the signature dishes — apprentice is faster but learned the essentials.

Scenario

Distill GPT-4o routing decisions into a 3B classifier — 90% routing accuracy at 1/100th inference cost.

Solution

In Model Engineering (Awareness), apply Distillation to this scenario: Distill GPT-4o routing decisions into a 3B classifier — 90% routing accuracy at 1/100th inference cost. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Do this before moving to the next module — reading alone is not enough.

Open the Code Walkthrough below and run it locally. Change one parameter related to Distillation (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Highlighted lines show where Distillation happens in the code.

1# Distillation — minimal example2from openai import OpenAI3
4client = OpenAI()  # create API client5
6# Ask the model to explain this topic7response = client.chat.completions.create(  # core API call for Distillation8    model="gpt-4o-mini",9    messages=[10        {"role": "system", "content": "You explain distillation clearly."},11        {"role": "user", "content": f"What is distillation?"},12    ],13    temperature=0,14)15print(response.choices[0].message.content)  # show output for debugging

• Treating Distillation as a black box without evaluation
• Ignoring cost and latency in production
• Skipping error handling for distillation

Quick recap — the most important points from this module.