Michele LaurelliBlog

Free tool · browser-only calculation

Softmax and temperature explorer

Adjust logits and temperature, inspect softmax probabilities and measure the entropy and confidence of the resulting distribution. With guided examples and.

How the calculation works

pi=ezi/T∑jezj/T,H(p)=−∑ipiln⁡pip_i=\frac{e^{z_i/T}}{\sum_j e^{z_j/T}},\quad H(p)=-\sum_i p_i\ln p_i
Formula as plain textp_i=\frac{e^{z_i/T}}{\sum_j e^{z_j/T}},\quad H(p)=-\sum_i p_i\ln p_i

Softmax turns a vector of scores into positive probabilities summing to one. Subtracting the maximum before exponentiation improves numerical stability without changing the result.

Temperature divides logits: low values concentrate mass on the most likely class, while high values make the distribution more uniform. Entropy measures this spread in nats and clarifies relative confidence.

The tool does not estimate calibration or accuracy: it only shows how a distribution changes for the supplied logits. Configurations stay in the browser and can be shared through the URL.