next-token distribution

The full probability distribution a language model produces over every possible next token, before any token is actually chosen.

The next-token distribution is the probability distribution $P(w_i \mid w_{<i})$ a language model outputs over its entire vocabulary at each generation step, produced by applying softmax to that step's logits. A sampling strategy (greedy, temperature, top-k, top-p) turns this distribution into one chosen token.

Examples

Given "The capital of France is", the next-token distribution might assign 0.92 to "Paris" and spread the rest thinly across thousands of other tokens.