Not able to train the engine to create my own chatbot

Hi, I am trying to create my own chat bot and teach it everything I know about a subject that GPT knows nothing about. Essentially to create a professor on a niche subject.

Looking at youtube videos I can see you used to be able to do this in the playground interface but it’s not possible any more.

Then when asking chat GPT how to do it and to give code examples, none of the code examples work because it gives you deprecated code.

Third attempt at doing this involved looking at the docs and seeing under “fine tuning” that now you are supposed to prepare a .jsonl file in the right format and do something like …

openai api fine_tunes.create -t mypreparedfile.jsonl -m davinci

… on the command line.

Well I did this with my question and answer pairs and then went back to playground and saw that I had a new option on the drop list to select my fine tuned chat bot but even when I select that and ask it a question from my list that is in the .jsonl file I never get the answer that I taught it. Changing the gravity setting to 0 or 1 makes no difference either.

What am I doing wrong?

Is prompt training and fine tuning different? If so should I be starting with some tool other than that command line tool and if so, what is it?

Ok so now I am super frustrated.

I have been trying for days to train the engine with both a question and an answer that I provide and it will not work at all.

Every single code example that chat gpt gives me results in an error because it gives deprecated code examples.

Then it says you have to use fine_tuned_only parameters to get the prescribed answer every time.

Then it says not to do that and to just train it with gravity of 0 and that fine_tuned_only parameters don’t exist any more and when that does not work it goes back to telling you to use fine_tuned_only parameters.

A look at the api docs on the open ai website do not even break down a full list of parameters for the openai.Completion.update method or tell you if you can use this call to do fine_tuned_only training.

Odd that an AI that knows how to code gives wrong answers every time you ask it how to code it.

Am pulling my hair out at this stage.

Has anyone trained the engine to give a prescribed answer to a specific question using the API? If so, please provide an example.

ChatGPT is meant for more general knowledge/information.

Have you checked out the Codex model that was trained on code?

Have you looked at some of the chatbot examples on GitHub?

1 Like