Top-P

Nucleus sampling — choosing from the smallest set of tokens whose probabilities sum to P.

A Top-P of 0.9 keeps the most probable tokens covering 90% of the mass.

Related terms

Back to Applications of Foundation Models