The impact of "presence_penalty" and "frequency_penalty" on "n" (many completions)?


When I ask GPT3 to generates many completions (i.e: n = 4), with “presence_penalty” and “frequency_penalty”. Does penalties have an impact on the next completions ?

Or penalty only look at the tokens given within the prompt ?


1 Like

I am curious about this, too.

Response from OpenAI support:

The penalty parameters applies to the context of each request within the given prompt.

1 Like

Thanks for coming back to share with us. Appreciate it.

1 Like

They are scenarios related to other scenarios that, in my opinion, do not apply to a simple “put 0.2” or “0.3” will generate more or less $this. after several months trying to validate our application I decided to start analyzing the contents generated in the last 2 months by analyzing the database that I have as a “save-backup”. Since I’m not an ultra mega power math lover, I managed to generate some Results using numpy, pandas , nltk and matplotlib. The ideal would be a continuous evaluation with Deep Learning to test scenarios different from what you are doing, in my case it would be fantastic to have time to study deep with summarization and a virtual friend .

In short: There is no manual or recipe for an iidea response.

Towards Data Science › …
GPT-3 Parameters and Prompt Design | by Anthony Cavin

To be sure, I asked for clarification, the given answer is:

The penalties do not have an impact on the next “n” completions.