Fine Tunned Model Doesn't answer like i want

API
1

Guys, I’m a newbie to fine-tuning, I started learning about it yesterday, and to test it I decided to create a model that would be trained with all the prices of products in a restaurant. However, when I ask questions in general he was very assertive with some products, but other products he simply invents a random price or even changes the name of one of the products even though I have already trained him with the price and put it for him to use the price that was spent on his training.

The training I put together, I asked 10 variations of questions for 6 menu items, my questions it’s like:

  • " How much does BigMac cost??"
  • “what is the price of bigmac?”
  • “How much is a Big Mac?”,
  • " What’s the current price for a Big Mac?"

Well you understand, and in the answer i do the same answer for all

-" Bigmac costs $5.69 " .

The way I trained him in the system saying that “you is an employee specialized in the menu and answer for costumers the price of the products in a assertive way”.

But for example some times he answer " Super bigmac costs $8.29" but i didn’t train he with this " Super Bigmac" cost.

Does anyone know if this is a lack of training? Do I need to add more data and more variations to improve efficiency? Or am I doing the training the wrong way?

2

Hi,

Welcome to the developer forum.

This is a common misunderstanding about fine tuning, it does not encode data. It encodes style. Best example I can give is, if you trained it on the works of an author, it would know very little about the books content, but it would be able to write another book in that same style.

If you wish to include prices and the list is not too long, I’d include them in the prompt, if you have a very large list of prices, tens of thousands, then I’d use a RAG pipeline to allow a vector search to pullback items that match a clients requirements and present them with the data that way, or you could pass that vector retrieval to the LLM model for further interactive processing with the user.

4 Likes
4

Hmm, I understand… but what I find strange is that when I compare the trained model with the untrained model, the trained model responds with exactly the values I input on train, only occasionally making a mistake. Doesn’t that mean it has ‘learned’?

5

Kind of, but that is not it’s primary purpose, it’s only very thinly learned and will often make mistakes. Hense the reason for using RAG.