temperature

A knob that reshapes a next-token distribution before sampling -- low temperature sharpens it toward the top choice, high temperature flattens it toward uniform.

Temperature $T$ rescales logits before softmax: $\text{softmax}(z / T)$ . As $T \to 0$ , sampling approaches greedy (always the highest-probability token); as $T$ grows, the distribution flattens toward uniform, making low-probability tokens more likely to be sampled.

Examples

Temperature 0 makes generation deterministic; temperature 1.0 samples directly from the model's own distribution.