Sampling only from the smallest set of tokens whose probabilities reach p.
Aucun attribut comparable n'a encore été enregistré pour ces deux technologies.
Généré à partir des différences ci-dessus par des règles structurées, et non par un modèle.
Choose Top-p Sampling if…
Aucun des attributs enregistrés ne le distingue à son avantage.
Choose Beam Search if…
Aucun des attributs enregistrés ne le distingue à son avantage.
La page reste accessible mais est exclue des index de recherche tant qu'elle ne contient pas assez de détails sourcés (PRD SEO-08).
technique · Techniques
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.
technique · Techniques
Beam search expands several partial sequences at once and keeps the highest-scoring ones, rather than committing to the token chosen at each step. It solves the problem that greedy decoding can be locally optimal and globally poor. It exists as the default for translation and summarisation, and it tends to produce bland text for open-ended generation.
