I’m trying to train a model using gpt-3.5-turbo-0613 to respond to a history of financial transactions that I made, when training with all the available data I had and several examples using the role and content model, after training when I call the chatCompletions api with my trained model, in questions more directed to text or even dates, such as what type of transaction it was and when, I get the correct answer I wanted, but every time it involves values and numbers it’s as if my model ignored the values it was trained on and created the values itself. As I’m new to using openai, I wanted to know:
Is it possible for me to train a model that can be assertive in the information I provided.
If it is not possible to train with fine-tune, is there a better way to do it?
These models are not good at math. It’s simply not a thing they’re built for.
If you need accurate data retrieval, use a traditional database.
If you need correct mathematical evaluation, you may be better off asking the model to generate some Python or JavaScript to solve the problem, which you can then execute.
Retrieval of accurate information is not what fine-tune is best used for. Your data becomes mixed with other data. It would take a special input prompt that makes a new identity, and a very high number of epochs, to make a model that is to specialized to do anything but repeat back only the format you trained on - and then you still might get hallucinated data that just “looks like” what it is supposed to look like.
You will likely want to use a semantic search database, that can retrieve parts of knowledge related to a query, and inject them directly into a prompt so the AI can answer.