Temperature
A sampling parameter that rescales token logits before probability normalization, controlling the sharpness and randomness of generated output.
Language-Model Context
Temperature rescales token logits before probability normalization. Lower values sharpen the distribution and tend to concentrate sampling on high-probability tokens; higher values flatten it and can increase diversity. The observed effect depends on the model and any additional sampling filters.
Decoding Boundary
Temperature does not add knowledge or change model parameters. It only transforms the probability distribution used during decoding.