Repair an overconfident model with temperature scaling

For yes/no probabilities, temperature scaling divides the logit by a single number $T$: $\hat p' = \sigma(\text{logit}(\hat p)/T)$. You fit $T$ by minimising log loss on a calibration split and evaluate on a separate test split, so the fit is never graded on the data it was fitted to. The search runs over $\log T$ rather than $T$, because $T$ is a scale: halving it matters as much as doubling it.