Top-p Sampling
Sampling only from the smallest set of tokens whose probabilities reach p.
Overview
Top-p, or nucleus sampling, keeps the most likely tokens until their cumulative probability reaches a threshold, then samples from that set. It solves the problem that a fixed top-k includes nonsense when the distribution is peaked and excludes reasonable options when it is flat. It exists because the set of plausible continuations should be sized by the model's own confidence.
Technology Graph
How Top-p Sampling connects to the rest of the stack.
Alternatives
-
Beam Search
technique
An alternative to Top-p Sampling · source
Comparisons
Side-by-side breakdowns with the alternatives above.