Test training Davinci and completion after training

I have created a Davinci FineTune model in GPT-3 with 2 sample prompts:

Prompt:

When was Company X started?

Completion:

Company X was started in 1984

Prompt:

How old is Company X?

Completion:

Company X was started in 1984

When I do a completion on that model with the exact same prompts, I don get the correct answers. I’ve tried different temperatures from 0 to 0.9. N = 1, Different token sizes, stop sign and so on.

But I don’t get the right answer. I am sure I am using my own model in the completion. I get results like Company X was founded in 2003 by a couple of brothers in Denmark. These completions have noting to do with the real answer.

Will it not work with only two prompts in the finetune model? GPT-3 says it will work better after 100 prompts, but they can’t mean that I should have like 100 prompts of the same question?

Shouldn’t I get the correct answer?

Welcome.

I don’t think 2 samples is nearly enough for fine-tuning. You probably want to start with 200 or more.

Fine-tuning won’t perfectly “memorize” a set of facts you feed it in the dataset.

Do some searching for “fine-tune” or “fine-tuning” here, and I’m sure you’ll find some related threads that might help you have a better idea of how it works.

Good luck!

2 Likes

If you have a limited number of examples, try increasing the n_epochs value. It will reinforce the learning significantly. The online help explains the values (The default is 4)

4 Likes