Neural Networks & LLM Components · Easy

Stable Softmax

Convert logits into probabilities without overflow or underflow.

Task

Implement stable_softmax(logits: list[float]) -> list[float]. Convert a one-dimensional sequence of logits into probabilities using the numerically stable softmax transformation.

Requirements

Subtract the maximum logit before exponentiation.
Return probabilities in the original input order.
Return an empty result of the selected implementation type for empty input.
The returned probabilities must sum to 1.0 for non-empty input.

Example

logits = [1.0, 2.0, 3.0]

stable_softmax(logits)
# [0.09003057, 0.24472847, 0.66524096]