Top-P
A sampling parameter for large language models that accumulates token probabilities in descending order until the cumulative probability reaches the threshold P, then samples only from that subset of tokens.
Core Facts
Timeline (last 90 days)
Temperature, Top-K and Top-P are the three core sampling parameters that control the randomness of text generation, with a common combination being temperature 0.7~0.8, Top-K 40 and Top-P 0.9~0.95.
All Facts (3)
Top-P (nucleus sampling) samples from the smallest set of tokens whose cumulative probability exceeds the threshold P, while Top-K selects only from the K candidates with the highest probabilities.
80%UnverifiedCoze's Top-P nucleus sampling mechanism starts from the highest-probability words and accumulates probabilities until the cumulative probability reaches the set P value, sampling only from those words.
60%UnverifiedTemperature, Top-K and Top-P are the three core sampling parameters that control the randomness of text generation, with a common combination being temperature 0.7~0.8, Top-K 40 and Top-P 0.9~0.95.
50%