Temperature

What Temperature is

Low temperature makes output focused and repeatable; high temperature increases variety and the chance of drift or error.

How it works

Temperature rescales the logits before sampling. Near zero the model almost always picks the highest-probability token. Top-p and top-k truncate the candidate pool as complementary controls.

Why it matters

Getting this wrong is a common cause of unreliable production behaviour: extraction and classification want near-zero, brainstorming wants higher.

Common uses

• →Deterministic extraction at low temperature
• →Creative ideation at higher values
• →Diverse sampling for self-consistency