Top-p sampling
Also: nucleus sampling, top-p
Keeping the smallest set of tokens whose probabilities add up to p, then sampling from that set. The cutoff moves with the model's confidence: a near-certain continuation leaves one candidate, an open-ended one leaves hundreds.
A definition is not an understanding. This is taught properly in How a model picks the next token, part of Sampling & decoding.
Learn this properlyTerms from the same topic