Frequency and Presence Penalties
Drafted replies had a tic: "Fantastic question! We'd be fantastic to help, and your bike will look fantastic." And a brainstorm of sale ideas kept circling the same two themes, discounts and free bells. Some APIs offer two settings aimed at exactly these problems.
The rule
Frequency punishes repeats by count; presence punishes any reuse once.
Both settings lower the score of tokens that have already appeared in the text, before the next token is chosen. They differ in how the penalty grows.
- Frequency penalty — the penalty grows with how many times a token has appeared. A word used four times is pushed down much harder than a word used once. It targets repetitive wording.
- Presence penalty — a flat, one-time penalty for having appeared at all. Once used, a token gets the same push down whether it appeared once or ten times. It nudges the model toward words, and therefore ideas, it has not touched yet.
Token "fantastic" used so far: 1 time 4 times
Frequency penalty 0.5: -0.5 -2.0
Presence penalty 0.5: -0.5 -0.5The numbers are subtracted from the token's score, not its final probability, so the effect on the odds is strong but not exactly proportional. Typical useful values are small, around 0.1 to 1. Not every API exposes these settings, and the exact formula can differ, but the frequency-versus-presence idea is common.
The shortcut
Name the problem first:
- "It keeps saying the same word" → frequency penalty.
- "It keeps coming back to the same topic" → presence penalty.
- "It repeats a required word, like the shop name" → no penalty; that repetition is correct.
Name problem one
"fantastic" three times in one reply: repeated wording. Choose a frequency penalty of 0.5.
The trap is turning penalties up high as a cure-all. Large values make the model avoid words it genuinely needs, and output grows strange. Often a clearer prompt ("use varied wording; do not repeat ideas") fixes the problem first.
Drills
Answer fast — trust the rule.
A token has appeared five times. Which penalty hits it harder than a token that appeared once?
A brainstorm keeps returning to the same few ideas. Which setting fits best?
Why can a very high penalty damage a Spoke & Chain reply?
With a presence penalty of 0.4, how much is a token penalised after appearing three times?
Remember
- Frequency punishes repeats by count; presence punishes any reuse once.
- Shortcut: same word → frequency; same topic → presence; required word → no penalty.
- Trap: cranking penalties up high instead of first fixing the prompt.