How do you structure a dataset like that? specifically conversational sciences. Assuming you use openai for finetuning where you need to follow a chat completions structure...