Tinker Small models: distillation & routing

by Kondasamy Jayaraman · drawn with Colorfigures by April Zhu

...

The bench before the release gate.

Chapter 6 of 7

Gate the switch

The router saves money at some thresholds and loses answers at others. Before it replaces the all-large assistant, the release gate from the Evals module gets a say.

This chapter: How do you prove a cheaper router is safe to ship?

You will learn

  • Why the all-large assistant is the baseline the router must match
  • How the gate’s two rules, a pass rate and no regressions, catch a confident wrong answer
  • How to find the cheapest setting that still ships
Words you’ll meet (3)
Baseline
The version a new one is measured against: here, every question sent to the large tier.
Regression
A question the baseline got right that the new version gets wrong.
Release gate
The rule that decides, from a test run, whether a candidate ships.

The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.

Step 1 of 4

Larkspur ships the all-large assistant today, so it is the baseline. Evals’ release gate has two rules: 90% must pass, and no question all-large gets right may fail. Run the baseline through it first.