Is there a way to restrict GPT-3’s response to only the text provided in Q&A feature?
Even if I keep the temperature =0, GPT-3 answers outside the text. GPT-3 should not respond based on its prior internet-based knowledge.
How to restrict the GPT-3 to its Q&A passage only?
I am using question-answering endpoint.
Often times GPT-3 is answering from outside the text file.
I want to restrict GPT-3 such that it doesn’t answer outside of the text file content.
This doesn’t work =>
examples_context = “Only use the below documents.
I am unable to understand below. Please provide example. Where do I get return_prompt?
Set the return_prompt = “true” look at the returned prompt and see if you could come up with a better one by making changes to the very beginning and/or submit the prompts to different models (instruct series).
You’ll also want to be sure you’re using an Instruct series model if you add an instruction like "Only use the below documents.” or “Answer ‘I don’t know’ if the question is unrelated to the text”. This has worked well in my tests, but it is admittedly a tricky problem to manage. A more in-depth option would be to fine-tune a discriminator: Playbook: Fine-tuning for Truthfulness [External] - Google Docs.