All terms
Inference & Serving
Repetition Penalty
A sampling control that discourages a model from repeating tokens it has already used.
Definition
Repetition penalty discourages a model from reusing words it has already produced by dividing the logits of previously generated tokens by a factor greater than one before sampling. This lowers their relative probability, reducing loops and redundant phrasing. Set too high, it can push the model away from useful words; set too low, the text may still repeat. It is commonly combined with temperature and top-p.