Active Nerds
Back to Glossary Index

Temperature (Sampling)

Hyperparameter scaling logit probabilities before softmax to tune generation randomness and entropy.

Mechanism & Definition

A sampling hyperparameter that scales raw output logits ($z_i / T$) prior to softmax probability distribution computation. Low values ($T \to 0$) produce deterministic, greedy selection of high-probability tokens; higher values ($T > 0.7$) flatten distributions to increase diversity and creative variance.

Related Terms

Appears In