Logit

A logit is a raw, unnormalized score output by a model before softmax (or similar) converts scores to probabilities.

Temperature scaling divides logits before softmax; cross-entropy training operates on logits via log-softmax for numerical stability.

← Back to AI Glossary