I am struggling to understand via the documentation how to fine tune. I think I do it in JSON format but then I just get confused from there. Does anyone know a course that walks through it step by step and is a site not full of ads and just trying to sell me something lol
Welcome to the community.
What is your use-case for fine-tuning?
I currently got the grant via this Microsoft for Startups and my use case is to train it with health data thus the problem this type of data would be huge but I have a few solutions to that. HOWEVER I am kinda a beginner here so I thought start small, I got basic question response via the API, easy stuff in Python, whats hard is know how to exactly fine tune it, I tried to talk it through with ChatGPT, but it gives varied results, to try and simplify it. I asked it if I had say 5 different fruits and there list of ingredients that make up the fruit how would I fine tune it, sometimes it gives me a vast JSON model, sometimes it gives me text, sometimes I have no idea what it is talking about. I tried to read through step by step on the docs this confuses me even more, it sounds like you do a JSON file of question and desired outcome? But I think overall how do you specifically even save the fine tuning and apply it your model that you are using?
Note I can show some example code if need be, I managed to make a build system in sublime text 4 that also scrapes websites (e.g. just extracts a bit of text nothing dodgy) and I have a python flask app I written that literally just sends the prompt and recieves an answer back via the API.
I feel like I am swimming through a vast ocean of snake oils salesmen (e.g. tutorial on youtube makes no sense and they are just selling something) with no actual guidance. Sorry if I seem a bit dumb here as well, my background is AI I done a degree in it even but we never really touched anything like this, in fact most my ideas so far are to mix in complex AI algorithms and such (but again I need to start small) I think my main obstacle at the moment is having a decent action plan in place to learn how to fine tune and move forwards effectively, is something is blatantly obvious I am not doing here please let me know, I have various learning difficulties… the documentation I don’t find too user friendly to this effect despite my skills in programming actually being relatively intermediate I think.
So an update I tried to reword what I am stuck on, I also posted this on Discord but its rare I get an answer on there:
OpenAI API I have tried to read this 1000 times, I don’t know if I am too tired. But anyway a bit of background I am trying to make a mental health support chat bot using Flask in Python. So the main language I am using is Python here. The problem is how do I actually store the fine tuning? I think I know how I do it, I convert my data from text / excel into JSON format in various ways, it accepts this? I need much help and guidance into steering my chat bot in the right direction as a web app. The web app side of programming I am fine with, its the fine tuning in between, people have just told me its too expensive to do, and I am like, yeah I know thats not my problem, my problem is how do I do it so I only have to do it once?
I’d recommend reading the fine-tuning docs, it’ll help you gain more clarity of the process.
Fine-tuning is to guide model’s “behavior”. Behavior would mean, how the model responds to particular completion requests.
To build your own fine-tuned model, you’ll need a JSONL file, where each line is a valid JSON object containing the prompt-completion pairs.
You should build a JSONL file with the required data, i.e prompts, and how you want the model to respond to particular to them.
You should make sure that the training data adheres to these guidelines.
Then you can upload the file and proceed with fine-tune process.
If it completes successfully, you’ll get a new fine-tuned model.