2
$\begingroup$

I am trying to fine tune a model using OpenAI's fine tuning API. I am passing bodies of text (for example, news paper articles) as prompts and the data I want from it as completions.

Let us consider the following: if a newspaper article I passed as a prompt to fine tune the data, consists of some information that GPT did not know before, like 'lithium ores are found in India'. If I use the completion API after that to ask GPT 'are lithium ores found in India?' as a prompt, will GPT be able to answer 'yes'?

$\endgroup$

1 Answer 1

0
$\begingroup$

Does GPT-3 remember data from prompts used to fine tune it?

Very unlikely, according the LIMA: Less Is More for Alignment paper:

These results strongly suggest that almost all knowledge in large language models is learned during pretraining, and only limited instruction tuning data is necessary to teach models to produce high quality output.

Also note that labels in prompts don't seem to matter much, see Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?.Sewon MinXinxi LyuAri HoltzmanMikel ArtetxeMike LewisHannaneh HajishirziLuke Zettlemoyer. EMNLP 2022.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.