Answers API Repetition in response

Hi,

I have been exploring the Answers API with a pre-uploaded document of FAQs in JSONL format.

It works very very well. But at times the completion gets stuck repeating itself until the MAX_TOKENS is reached. Is there a good way to stop this happening?

On the main completion API there is a frequency_penalty and presence_penalty to reduce this but is this not supported on /answers?

Thanks
Dave

EXAMPLE COMPLETION:
The solution provides a role-based access control model that allows users to be assigned roles that define their access to the platform. Users can be assigned roles that define their access to the platform. Users can be assigned roles that define their access to the platform. Users can be assigned roles that define their access to the platform. Users can be assigned roles that define their access to the platform. Users can be assigned roles that define their access to the platform. Users can be assigned roles that define their access to the platform. Users can be assigned roles that define their access to the platform. Users can be assigned roles that define their access to the platform. Users can be assigned roles that define their access to the platform. Users can be assigned roles that define their access to the platform. Users can be assigned roles that define their access

1 Like

Some recommendations:

  • Decrease response length (to make that amount of repetition impossible)
  • Increase temperature (to increase randomness, and therefore decrease repetition)
  • If you’re not already using the highest-end completion model (davinci or davinci-instruct-beta), I’d try that as well
  • Ensure that there are varied examples with no amount of repetition
  • Separately use Search + Completion to control the penalty parameters
1 Like