Trying to fine tune davinci to know new objects

Community
1

Hey there!
I’m relatively new to fine tuning gpt-3 models therefore any tips would be amazing!

I created a few different robots and I wanted to let gpt-3 explain them to someone when they are asking questions about a specific robot. In the following text I will just refer to the robots as objects.

I tried to fine tune a davinci and a curie model to know and understand new objects.
(The non-finetuned models have no knowledge about them)

I have about 15 to 20 different new objects.
How can I bring davinci or curie to the point that they know the objects?

I tried it like this:

  1. Creating a new dataset with information about the objects.
    like:
  • What is object1 ?
  • What does object1 do?
  • What is special about object1?
  • What is object2?
  • What does object2 do?
  • What is special about object2?
    …
  1. Writing responses to the questions above
  2. Fine tuning

What I noticed is that both models Curie and Davinci are sometimes mixing up information.
Generating information of object2 and putting those into the text when I ask for object1.

Is there a way to make it more consistent and “bind” them more to the objects?

2

fine-tuning is not the best approach to add knowledge

3

what would you suggest?

4

search using embeddings, then use the result in the prompt to generate a final response to the question

3 Likes