MODEL INTERFACE / SAMPLING

Temperature and next-token probabilities

Hold the five logits from a fixed, untrained model constant. Move the slider to change positive temperature τ.

pᵢ(τ) = exp(zᵢ / τ) / Σⱼ exp(zⱼ / τ)
1.00
0.203.00
b / ID 1Original logit 2.0967Scaled logit 2.0967

Bars use full-precision probabilities; labels round to two decimals. Positive temperature preserves the ranking: greedy still selects b / ID 1.