Tinker Synthetic data

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

“Q: Does the Kestrel K3 come with a charger? A:”

Chapter 2 of 7

Ask the model to write

The team’s idea: let Larkspur mini write more examples. Synthetic data, written by a model, not by people.

This chapter: What does a model write when you ask it for twenty replies?

About this chapter

You will learn

  • How temperature trades repeats for variety
  • Why a cool model writes the same reply again
  • Why a hot model drifts off topic

Words you’ll meet

Synthetic data
Training examples a model writes, instead of people.
Temperature
How adventurous sampling is: low picks likely words, high takes more chances.
Copy
A reply that repeats, word for word or mostly, one already written.

Step 1 of 4

is text a model writes. sets how adventurous each word pick is: low plays safe.

Predict first.

Twenty replies at temperature 0.7, then twenty at 1.3. Which batch names Larkspur’s products more often?