...
Max Tokens: Limits output length; consider a token 75% word.
Usually between 1 and 4096 tokens.
For example, set to 130 for brief answers, about 100 words, and 800 for long-form content, like 600 words.
Michael says multiplying your desired word count by 1.33 gets a good token value.
Temperature: Controls randomness, AKA creativity.
Typically from 0 to 1.
E.g., set to 0.2 for more deterministic outputs and 0.7 for creative brainstorming.
Top-p: Controls output diversity by limiting the scope of the related data used.
Generally, from 0 to 1.
E.g., set to 0.9 for more varied responses and 0.5 for focused, narrow outputs.
Frequency Penalty:Deters Influences repetition of similar words or phrases.
E.g., set to Typically from -2 to 2.
Negative values, such as -0.5 to avoid recurring phrases, 0 for natural conversation flow, and over 0 for more repetition.
Presence Penalty: Encourages novelty, like using rarer words.
E.g., set to 0.6 to promote new concepts and 0.9 for groundbreaking ideas25, lead to increased repetition. Beneficial where consistent terminology is crucial or when emphasizing specific points or concepts is desired.
0.1 attempts to avoid repetition, yet it may still use the same word or phrase again when relevant.
0.9 actively avoids repeating words or phrases, often using synonyms or rephrasing sentences to ensure variety.
Presence Penalty: Introduces new topics or ideas in its responses.
Usually ranges from -2 to 2.
At 0.6, actively encourages introducing new topics or ideas, often shifting away from the current.
At 0.1, slightly encourages new topics or ideas, resulting in less pronounced shifts.
At -0.1, gently discourages new topics and stays closely aligned with minimal deviations.
At -0.6, strongly discourages new topic introductions, keeping the conversation focused tightly.
📝 Create ChatGPT Settings based on the persona or task.
...