Hello, I am trying to train my assistant on platform.openai, so that I can integrate it on my website. I have Written instructions and give it 3 MB worth of JSON file. Basically it is history of different parties that go to elections. Thing is that no matter how many instructions I give few problems always arises:
- Answers are inconsistent , sometimes assistant chooses correct data from the json, sometimes it gives false facts as information
- It always gives too short answers
- Cannot make it make references. In json there is type : “references” and I want to make assistant read links from this, but instead it generates link to data.json itself, the file which users cannot access
Maybe you know how to fine-tune this or maybe I should use different methods for training
Welcome to the community!
You used the fine-tuning tag but it sounds like you’re just using files?
Do you have more details? Some code and your prompt?
I am quite new to this, maybe I use terminology wrong here.
So main goal here is to take data about elections and base all the answers from this data, which it does but too briefly and sometimes inaccurately, if it is not possible doing it by only instructions and provided data, could you suggest any plan or resources as for how to train model better?
2 Likes
Yeah, fine-tuning is more for “style” rather than “facts” or “knowledge”…
I would look into Assistants API…
There’s an overview in the Cookbook…
Or if you have more specific questions, let us know! 
2 Likes
Only specific question that comes to mind - is there some way except maybe structuring my data better so that Assistant can read from it more accurately and won’t damage the facts I provide. (Maybe it cannot process all this better since it is not in English but Georgian and 4o struggles with it).
Thanks for your replies btw