AI Glossary

Frequency and presence penalties

Also called repetition penalty

What does “Frequency and presence penalties” mean?

Frequency and presence penalties: Settings, offered by some APIs, that discourage a model from repeating itself. A frequency penalty grows stronger the more often a token has already appeared; a presence penalty applies once a token has appeared at all, nudging the model toward new words and topics. Strong penalties can make text read unnaturally.

Frequency and presence penalties visual guide: Settings, offered by some APIs, that discourage a model from repeating itself. A frequency penalty grows stronger the more often a token has already appeared; a presence penalty applies once a token has appeared at all, nudging the model toward new words and topics. Strong penalties can make text read unnaturally.
Visual guide Open full size (new tab)

What is an example of “Frequency and presence penalties”?

A small presence penalty can stop a product-description generator from reusing the same adjectives in every line.

Keep learning

See also

TemperatureTermTop-pTermTokenTerm

More from Franklin

On Franklin

Model parameter calculatorPage

Same topic

GroundingTermFew-shot promptingTermHallucinationTermEmbeddingsTermIn-context learningTerm