Hi, i’m trying to perform a fine-tuning but i’m facing issues.
I created the jsonl file using the tool:
openai tools fine_tunes.prepare_data
The file now looks as follows:
{“prompt”:“What is a performative space ->”,“completion”:" A performative space is a space with theatrical qualities\n"}
{“prompt”:“How can one create a performative space ->”,“completion”:" A performative space can be created by putting together a composition of surfaces that allows for a theatrical procession\n"}
{“prompt”:“How do surfaces create a spatial environment ->”,“completion”:" Surfaces come together to create a spatial environment by formulating a roof, walls and a floor\n"}
So i accepted all recommendations:
- Add a suffix separator
->
to all prompts - Add a suffix ending
\n
to all completions - Add a whitespace character to the beginning of the completion
I launched the fine-tuning and now when i try to test it i’m facing issues. For example in the playground if i choose this new model and type:
hello how are you? →
and then i click submit, this is what i get:
hello how are you? → Hi John my name is John
John Hi my name is John how are you?
John Hi my name is John how are you? → I’m good thank you and you?
John Hi my name is John how are you? I’m good thank you and you? → Me too
John Hi my name is John how are you? I’m good thank you and you? Me too
John Hi my name is John how are you? I’m good thank you and you? Me too → What brings you here?
John Hi my name is John how are you? I’m good thank you and you? Me too What brings you here? → I’m here for a haircut
John Hi my name is John how are you? I’m good thank you and you? Me too What brings you here? I’m here for a haircut → A haircut for a man or a haircut for a woman?
John Hi my name is John how are you?
This happens for every prompt i write. I also tried using the APIs and the result is the same
What’s wrong with it? Am i using the fine-tuning in a not proper way?
Thanks in advance
Stefano