I’m using gpt-4o + RAG to generate documents as they’re made in my company. It’s working well but I need it to achieve better results, that’s why I’m finetuning it and these are the kind of entries I’m writing in my .jsonl:
When model replied correctly with a well made document:
{“role”: “user”, “content”: “Make me a document within this info”},{“role”: “assistant”, “content”: “Right document”}
When my model fails:
{"role": "user", "content": "Make me a document within this info ..."},{"role": "assistant", "content": "Wrong document"},{"role": "user", "content": "You failed in the following points: .... Make it again"},{"role": "assistant", "content": "Right document"}
Hi, too little background to say something meaningful. Out of the box, it looks like the task is too complex to handle in one run. A better design of the workflow would probably help, but as I said, too little info (input size, format, type of doc, format, length, complexity, etc) to help.
Sorry, I’m actually looking for a more generic solution. Regardless of the task, I just want to confirm if this is the correct way to perform a fine-tuning job, or if the ‘I reply when it’s wrong and let it be when it’s okay’ approach is flawed Thank you for replying!
As a more general approach, all of those tools are good: RAG, fine-tuning, assistants, coding…
The question is are they adapted to your specific goal? And here the goal is not clear. Because from the little snippets I see, for me, they don’t even fit into one step, so fine-tuning here would basically break your application. But then without the additional information about what are you trying to achieve it’s like writing on the water with a stick.
Thank you for replying!