ClusterChat

Joining…

The model

—

Settings

Sampling

Temperature0 = greedy argmax Top-kSample from the k likeliest tokens. 0 = off Repetition1.0 = off Frequency penaltyPer repeat of a token already seen. 0 = off Presence penaltyOnce for any token already seen. 0 = off No-repeat n-gramForbid repeating an n-gram. 0 = off Max tokens0 = until the reply ends or the context is full

Prompt

ModeHow the prompt is fed to the model

Diagnostics

DecodingWhere each next token is sent from Speculative decodingCheck guessed tokens several at a time ProfileTime each token on every peer it crosses; shown under the reply