Glossary

Temperature

Beginner

A knob that makes a model’s word choices safer (low) or wilder (high).

Novice

A setting that sharpens or flattens the next-token probabilities before one is picked. Low temperature nearly always picks the top token; high temperature spreads the chances out.

Expert

Logits are divided by TT before softmax: T→0T \to 0 approaches greedy decoding, T=1T = 1 is the model’s distribution, T>1T > 1 flattens it.