I’m trying to fine-tune a model on text data from 3 books for text completion jobs. Previously I have been able to do this using black “prompts” in the jsonl file, with paragraphs used as the “completion” text. When I try to do this through the API now, I get an error message relating to the empty prompt fields. Is there a way around this please?
Still works fine for me fine tuned a model last night with it in this format:
{"messages": [{"role": "system", "content": "My system message"}, {"role": "user", "content": ""}, {"role": "assistant", "content": "My completion"}]}
I kind of wonder if there’s really any difference in the type of reinforcement learning being done when you train on prompt vs completion, especially on a completion model.
For example, training without any separator on davinci-002 or babbage-002 is somewhat similar to the original pretraining on books, wikipedia. In the base model you can have it complete on random text and you might have the form of real estate listings or reddit posts continued after the text. The pretraining both can be activation text and output text. Does it make any difference if you stick a period character as the only prompt or the only completion in your training file?
About the only thing that seems logical is that you don’t want to have a context training that is half Poe and then half Twain, because that might form a connection that one brings about the other.
So I would say: - add the period of the last prompt sentence to “completion”. - add just a space character as prompt. - train both cases, and see if they aren’t the same.