Frequency penalty
A setting that pushes down a word’s odds the more times it has already appeared.
Think of it like
Like a game where each repeat of a word costs you more points than the last.
Example
Generating product tags, a dev raises the frequency penalty so common words like “premium” stop dominating every list.
How it actually works
Unlike a flat repetition penalty, this one scales with how often a token has shown up — the tenth “the” is barely touched, but the tenth “premium” gets hit hard. It curbs runaway repetition while leaving genuinely common words mostly alone.
For product teams
Tune it to keep outputs varied when you’re generating lists or many short items.
For engineers
Subtracts a term proportional to a token’s prior count in the generation from its logit before softmax.
Related
- Presence penalty — Penalizes appearance regardless of count.
- Repetition penalty — The blunter version.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome