Hi,
I followed the process here to fine tune a davinci model with two sample entries:
{“prompt”:“Who is the owner of Jixdfdd company? ->”,“completion”:" kevin\n"}
{“prompt”:“Who is the owner of abcjdew company? ->”,“completion”:" bob\n"}
However, when I ask the question exactly the same with prompt, like “Who is the owner of Jixdfdd company? ->”, it doesn’t answer the trained completion “kevin”. I tried different temperature, but it still has no effect. May I know if I missed something?
Two sample entries are not going to sufficient enough to override the billions of parameters of training that the base model already has.
Fine-tuning is most useful for teaching the model a pattern or structure to respond in, not to train it exactly how to respond to specific questions.
If you really wanted to make this work, I would add a lot of variations to the question with the same answer for each one, so that the model can continue to reinforce this relationship. You could also increase the number of epochs for training, but this may lead to overfitting. It may also help to bump up the prompt_loss_weight parameter when fine-tuning from 0.01 default to 0.05 or 0.1, since the model needs to learn some of what is in the prompt to make the Q&A thing work.
Overall, though, I wouldn’t recommend using fine-tuning to try to teach the model new information. Try to find use cases for your fine-tuning where you are leveraging the vast amount of information the model already has about subject matter, and focus on formatting it in a useful way.