Invalid fine tuning training file even with a 34 character file that validates

So I kept trying to upload files for fine tuning and getting the Invalid File error, so after many frustrating iterations I gradually pared down my file to the following:

{“prompt”: “A”, “completion”: “B”}

That validates with the validator at JSON Lines, it is in UTF-8, and it was produced on Notepad++ on windows with Linux EOF characters.

Does anyone have any suggestions? I’m at the end of my rope, and especially wondering how a company full of coders can have their software generate such an unhelpful error message.

Any assistance is appreciated.

1 - you need 10+ examples
2 - you can only use “prompt” on a base completion model such as davinci-002
3 - if you are using completions, you’ll want to have a separator as “prompt” and a stop sequence, otherwise you’ve just trained the AI on “AB” and then who knows if it then gives you the training you did on a B token or goes into a loop.

I’ll bet you are trying to train gpt-3.5-turbo, though. In that case, you must construct example conversations in the form system/user/assistant messages just as when using the endpoint, where the system message is an identity that significantly departs from “You are ChatGPT”, and assistant is the behavior you want from such a user input. You’ll also need training coverage of every valid input scenario you might see, even the denials of going out-of-domain.

Yup, I re-checked things, and my problem was using ChatGPT. as a tutor. As you suspected it gave me the wrong formats for the model I was using. I got it to validate on davinci.

It’s still inconceivable to me that they don’t give a single bit of information on why these jobs fail.

Thanks again!