Finetuning question

Hey all!

I’m looking to finetune GPT to do a task for me with very little to no random deviations in its responses. I have 100 prompts and responses ready to finetune with, but if I train on so few examples, will I still get the pesky random responses? Is it worth using GPT to generate more responses and hand pick them to add to the training set?

Thanks in advance!

Yes, since you’ll be training a base model, you’re likely to get random aberrations like repetition. 200 samples is the bare minimum.

1 Like