TinkerLoading Larkspur mini
Chapter 4 of 7
One dial
The team likes the fine-tune’s style and won’t retrain it. Can its confidence be repaired after the fact instead?
This chapter: Can one number make an overconfident model honest again?
About this chapter
You will learn
- How temperature softens or sharpens every confidence
- Why the top guess never changes
- Why you set it on one half and check it on the other
Words you’ll meet
- Temperature
- A number the model’s scores are divided by before they become probabilities. Above 1 softens them.
- Temperature scaling
- Fixing a model’s confidence by choosing one temperature on held-out text, without retraining it.
- Log loss
- How surprised the model was by the real next piece, on average. Lower is better; it punishes sure mistakes hard.
Step 1 of 4
: the model’s scores are divided by it before they become probabilities. Above 1, every confidence softens.
Predict first.