I’m using a fine-tune model in which I’ve put different responses that I’d like my model to be able to respond to when I start it up. So far so exotic.
The problem is that the answer I get contains “12 or 31” or “Null” or even sentences that make little sense and some parts of the sentence may be missing.
Do you know how to remedy this ?
Is it a problem in my original jsonl model ? Knowing that I always use the same sentence in the prompt and to separate “prompt” and “completion” I use this “\n\n###\n\n”, is it important to change it or does it not matter ?