Fine-Tuning with Custom Data

Was trying to fine tune the gpt turbo 1106 model on my custom data, I have data from the Financial Report of Tesla for 2021 and 2022 where I am providing a set of questions along with the answers but after fine tuning the model when I am trying to query the fine-tuned model with the questions it is not giving the correct response, there is always an incorrect number.
Here is an example of my dataset -:
{“messages”: [
{“role”: “system”, “content”: “You are a chatbot who is also an expert financial analyst.”}, {“role”: “user”, “content”: “What was the total revenue of Tesla in 2022”},
{“role”: “assistant”, “content”: “In 2021, Tesla’s total revenues were $81.46 billion.”}]}

Instead of providing the number I gave in response it is providing some random number.

Can anyone tell me What I am doing wrong also the Hyperparameters for fine-tuning is auto and have around 113 data in the dataset of same kind.

Hi! Welcome to the forum!

Trying to use fine-tuning bake data into the model isn’t the best idea. It only really works if you want to tweak how it responds, not what it responds.

You may need to look at other processes, such as RAG for example, or maybe structuring your documents and using SQL or something to query them.

Also, I’d draw your attention to this, if you’re struggling with random results: