Unlike the frequency penalty, which penalizes more strongly as frequency rises, the presence penalty already takes effect on a word's first appearance. It encourages the model to go beyond already-mentioned terms and bring in new aspects. This is useful for brainstorming or when broad topical coverage is desired. Like the frequency penalty, it is usually set between 0 and 2.
Presence Penalty
Presence penalty is a parameter that down-weights words as soon as they have appeared even once in the output, to encourage new topics. Higher values lead to more topically varied text.