Fine tuning support

[elfin niche]

I have fine tuned a GPT3 model with over 5K training examples but I am not being able to get the correct answer from the trained model.

They all have a different library id and file id. Basically I am trying to show GPT what is the library id and file id for a set of subjects.

Each training example looks like this:
{"messages":[{"role":"system","content":"This chatbot provides the correct library id and file id based on the question."},{"role":"user","content":"What's the library id and file id for a media related to 'notification, picture, social, media, photography, photo, gallery, network' in the category 'illustrations'?"},{"role":"assistant","content":"The library id is 'lib_FepZzyPrOkHcFKf' and the file id is 'twd2vpqxsbytv3dl.png'."}]}

However when I ask a question to the trained model, I get a made up file id instead of an id I provided. Any suggestions or comments on why might not be working?

LD;DR

Training process:
notification, picture, social, media, photography, photo, gallery, network -> lib_FepZzyPrOkHcFKf/twd2vpqxsbytv3dl.png

Prompt after trained:
What would be the best suitable library and file for the sentence: Get messages instantly on your phone

Expected result:
lib_FepZzyPrOkHcFKf twd2vpqxsbytv3dl.png

[tight sphinx]

fine tuning is not at all optimal for storing precies ids / data. Please see https://platform.openai.com/docs/guides/fine-tuning/when-to-use-fine-tuning

[elfin niche]

@tight sphinx I thought with a lot training data that would be possible. I can see this being really useful for cases like mine. If you have any suggestions please let me know

the reason GPT could be great on this is because the prompt after trained might not be the same keywords from training process, but has to be inferred from the prompt

[tight sphinx]

This would be really hard to achieve with an LLM. You would need tons of data. In order if around 1000 more or more than what you have rn, and still idk if it would work. Also quick note, system messages work better at the second person (you are CHATGPT...)

That's why ai is extremely bad at giving back links. It's like if you asked an ai a YouTube video url. It wouldn't work unless it's a veeeeeeeerY famous video, and only sometimes.

The LLM in your case understands that it should answer gibberish, because there is no sense at a language level for that file name.

[elfin niche]

I see, thank you for your reply and explanation Paillat