logit bias
In plain wordsA setting that lets developers make specific words more or less likely in an AI model's output — even ban a word outright — by adjusting the model's scores before it chooses.
A decodingdecoding, n.: The method a language model uses to turn its predictions into actual text — for example, always picking the most likely next word, or choosing somewhat at random among the likely ones. control that adds a fixed, user-specified offset to the logitslogit, n.: A raw score that a model gives each possible answer before the scores are turned into probabilities. In statistics, the 'log-odds': a probability rewritten on a scale running from minus to plus infinity. of particular tokenstoken, n.: The small chunk of text an AI language model reads and writes — often a word, part of a word or a punctuation mark. AI usage, limits and prices are counted in tokens. before the softmaxsoftmax, n.: A formula that turns a list of raw scores into probabilities — all positive and adding up to one — with the biggest share going to the highest score., making them more or less likely at every generation step. Large negative values effectively ban tokens; large positive values all but force them, which can restrict a model to a fixed set of answer tokens for classification. In OpenAI's API the logit_bias parameter maps token IDs to values from −100 to 100, and other APIs and open-source engines offer similar options. Because it operates on token IDs, using it requires knowing how the tokenizertokenizer, n.: The component that splits text into tokens and converts them to numbers before an AI model sees it, then turns the model's output numbers back into text. splits the target words, including variants with leading spaces and capitals; banning a token does not ban the idea, which can surface through other tokens. It is a simple, model-agnostic precursor of fuller constrained decodingconstrained decoding, n.: Forcing an AI model's output to follow a required format — such as valid code or a fill-in form — by blocking, at every step, any word that would break the rules..
We set a logit bias of −100 on the em-dash tokens, and the model switched to commas.
Origin
Exposed as the logit_bias parameter of OpenAI's GPT-3-era completions API and later adopted by other providers and inference libraries; adding prior offsets to classifier scores is a much older technique.
Related
- logitlogit, n.: A raw score that a model gives each possible answer before the scores are turned into probabilities. In statistics, the 'log-odds': a probability rewritten on a scale running from minus to plus infinity.
- constrained decodingconstrained decoding, n.: Forcing an AI model's output to follow a required format — such as valid code or a fill-in form — by blocking, at every step, any word that would break the rules.
- repetition penaltyrepetition penalty, n.: A setting that discourages an AI model from repeating itself by lowering the chances of words it has already used, so its text does not get stuck in loops.
- tokenizertokenizer, n.: The component that splits text into tokens and converts them to numbers before an AI model sees it, then turns the model's output numbers back into text.
- decodingdecoding, n.: The method a language model uses to turn its predictions into actual text — for example, always picking the most likely next word, or choosing somewhat at random among the likely ones.
Mentioned in
biasbias, n.: A word with several AI meanings: an adjustable number added inside a model; a systematic error in estimates; or unfair treatment of some groups of people by an AI system., constrained decodingconstrained decoding, n.: Forcing an AI model's output to follow a required format — such as valid code or a fill-in form — by blocking, at every step, any word that would break the rules., decodingdecoding, n.: The method a language model uses to turn its predictions into actual text — for example, always picking the most likely next word, or choosing somewhat at random among the likely ones., logitlogit, n.: A raw score that a model gives each possible answer before the scores are turned into probabilities. In statistics, the 'log-odds': a probability rewritten on a scale running from minus to plus infinity., repetition penaltyrepetition penalty, n.: A setting that discourages an AI model from repeating itself by lowering the chances of words it has already used, so its text does not get stuck in loops., steeringsteering, n.: Guiding what an AI model does — its topic, tone or behavior — while it runs, by changing its instructions, its internal signals or how its next word is chosen, rather than by retraining it.