Definition
Temperature
Temperature is the randomness dial applied when picking the next token from the model's probability distribution. Near zero, the model almost always chooses the most likely token, giving deterministic output suited to code, extraction, and factual questions. Around 0.7 to 1.0 it explores more varied options for creative work. Above 1.5 output degrades into noise. Production settings usually combine it with top-p sampling.
Explained in
Chapter 5: Inside the Machine
Everything you need to know about what's happening inside the box. No PhD required.
Related terms