The best way to make a chat bot only answee question based on a given paper

[dominic.hofilena]

I currently have a research paper, https://cancerpreventionresearch.aacrjournals.org/content/3/1/108 which has information about abstract, method of study, experiments and conclusion about pomegranates and it’s cancer-fighting effects.

But its a long paper and will not fit a prompt. So I believe the best way to tackle this is through Fine-tuning. However, after testing the fine-tuned model. When I ask it a question, it would answer me with content that is outside the research paper.

How do I train or what is the correct prompt to use so that when I ask the model a question. It would only give me answers that is within the research paper.

ex: What is the purpose of this study?
and it will answer:
The purpose of this study was to investigate the antiaromatase activity and inhibition of testosterone-induced breast cancer cell proliferation by ET-derived compounds isolated from pomegranates.

Thanks in advance for anyone who answers

[daveshapautomator]

The INSTRUCT series works really well. I generally follow the format:

Read the following article and answer questions

[insert article]

Questions:

1. Blah blah blah?

etc.

[dominic.hofilena]

First. thanks daveshapautomator for checking this question out, really appreciate

So the INSTRUCT series would answer the question based on the overall “Mind” / “Wikipedia” information that Open AI has. and that’s a large scope of info

I was hoping to create a fine-tuned model so that the AI will only find its answer within the uploaded research paper.( https://cancerpreventionresearch.aacrjournals.org/content/3/1/108 )

and not go to far ahead.

[boris]

There’s also an /answers endpoint, which can be used when you can’t fit the entire article within the token limit.
This works by a two step process (but is handled on our end, within an answers endpoint):

1. find the most relevant part of the article.
2. answer the question based on that part of the article, as dave pointed above.

[boris]

Fine-tuning on a single paper isn’t meant to work. For fine-tuning you usually require at least ~200 examples - one paper is too little.

[dominic.hofilena]

Ohh there we go! Cool cool! OpenAI API. Thank you very much to you guys ( boris and daveshapautomator ). I guess I should play around with this.

Quick question Boris. If I add the whole Abstract on a single JSON line.

And ask a question, “What is the purpose of this study?”
Would it attempt to answer through a sentence? or it would return the whole abstract?

I guess that just something I have to test

[boris]

You can give a few example context & question & answer in the /answers endpoint, which will help the model know what kind of format you expect for answers.