MikeDoes
commented
2 years ago Added a pre-processing function, changed the sentence prefix to: "In the sentence", and fact prefix to "The facts are". Changed facts from array to triples (with parentheses) and removed quotations. When querieng, exclusing "The facts are" helped. Experimented with frequence and prescence penalties. Negative values did not help.
Test data was sent to GPT-3 Advice from Advisor:
If you want better performance with GPT-3 you will need to spend time modifying the prompt and the way you give the examples to the model to in order to make the problem description as natural as possible. -- In general, the better the problem is described (and the easier it is for a human to understand what the response should be), the better GPT-3 will perform.
Finally, one of the papers we discussed last time already performed one-shot learning with GPT-3, so you might need to think about what the novelty of your approach is....
Awaiting response from OpenAPI support