Hello everyone, my question is because I would like to fine-tune the model I use. I explain the process:
- I send a prompt with an attached document (PDF)
- I get a structured output from the Assistant based on the content of the document
- I send a true response to end the execution.
My point is the following: There are times when the structured output contains incorrect data based on the recognition performed on the PDF document, I could form a response with the same form as the structured output with the correct values to tell the model in some way: “Look, this is the PDF document you received, this was your response, however, this is the correct response.”
Is this possible? If so, what form should the .jsonl file have to be able to perform this task?