Tinker Calibration

by Samy (Kondasamy Jayaraman) · drawn with Colorfigures by April Zhu

Larkspur mini, ready to guess the next piece of text it never trained on.

Chapter 2 of 7

Honest jars

Before trusting the assistant’s numbers, the team checks a model whose every guess can be counted: Larkspur mini, this site’s tiny model.

This chapter: Is a tiny model honest about how sure it is of its own guesses?

About this chapter

You will learn

  • How a model’s odds for the next piece are a confidence
  • Why held-out text is the fair test
  • How to read a reliability diagram

Words you’ll meet

Top guess
The next piece of text the model gives the highest probability.
Held-out text
Documents the model never trained on, kept aside to test it fairly.
Reliability diagram
A chart of stated confidence against how often the answers were right. Honest models sit on its diagonal.

Step 1 of 4

Larkspur mini gives every possible next piece a probability. Its , and the probability it gave, are one marble.

Predict first.

When Larkspur mini puts 90% or more on its top guess, how often is that guess right?