Hello! Non-technical, but I’ve figured out how to pass a JSONL fine-tuning file to gpt-3.5 for fine-tuning.
I can wrap my brain around basic Q&A forms of fine-tuning using the conversational format (e.g. {“messages”: [{“role”: “system”, “content”…) , but I’m struggling with my use case.
As a journalist, if I take several of my articles (let’s say 2000 words each), how would I chunk those into this conversational format (system, user, assistant content) to impact the tone of newly fully AI drafted articles on some other topic of interview answers I’ve collected and fed to GPT to write the article on in the assistant/user/system roles? System is probably easy …you are a journalist, then how to handle the other two to achieve my end goal?
Fine tuning should reflect the types of inputs you will provide, and the types of output you expect from them.
After fine-tuning on a task, “help me rebuild my carburetor” won’t sound like “write an article in the nick.h style about carburetor repair” - if you’ve done fine tune right and not wasted your resources.
gpt-3.5-turbo has lots of training, including responding to and following different system prompts. You get that automatically. A good technique is to go in a new direction with an identity all your own for your fine-tune. System: “You are NickBot, a Nick writing style emulator”? AI is not going to be trained how to follow that already!
Then there is the fine-tune training data, and construction of example conversaations. From your unaltered sources such as articles, you’d put in a user instruction: “As NickBot, write an article on flower arranging that will appear in vanity fair” → your article. Tons of those, and then you can replicate that command and the AI will infer different outputs.
You might have other types of outputs that need more example data you don’t have if you have different jobs for the AI in your style.