← all blocks
calibration · v1.0.0

Temperature Scaling

Single-parameter calibration: divides logits by T. T>1 softens overconfidence; T<1 sharpens.

Contract

Input:
Predictions
Output:
Predictions
Determinism:
pure
Side effects:
none
Leakage window:
0s

Params (live form)

JSON
{
  "T": 1
}

Examples

  • Pull confident preds toward 0.5
Try in builder

Citations

  • Guo et al. 2017 — On Calibration of Modern Neural Networks