How do I include structured data in fine tuning dataset?

API Bugs
, , ,
1

How do i include structured data about businesses in fine tuning dataset. Example:

[
    {
        "id": 1,
        "name": "Test 1",
        "rent": "$10,000",
        "asking price": "$100,100"
    },
    {
        "id": 2,
        "name": "Test 2",
        "rent": "$9,000",
        "asking price": "$98,100"
    },
    {
        "id": 3,
        "name": "Test 3",
        "rent": "$10,000",
        "asking price": "$100,100"
    }
]

This is what i’ve tried:

First fine tuning dataset

{
    "messages": [
        {
            "role": "system",
            "content": "You are a helpful assistant that provides business information and set appointment."
        },
        {
            "role": "user",
            "content": "Tell me about business with id 1"
        },
        {
            "role": "assistant",
            "content": "id: 1\nName: Test 1\nRent: $10,000\nAsking price: $100,100"
        }
    ]
}

Second fine tuning dataset

{
    "messages": [
        {
            "role": "system",
            "content": "You are a helpful assistant that provides business information and set appointment."
        },
        {
            "role": "user",
            "content": "Business with id 1 rent"
        },
        {
            "role": "assistant",
            "content": "$10,000"
        },
        {
            "role": "user",
            "content": "Business with id 1 asking price"
        },
        {
            "role": "assistant",
            "content": "$100,100"
        }
    ]
}

I’ve uploaded these two different datasets and trained/fine-tuned them on the same model but when i ask the model: what is the asking price for business with id 1? It will give me wild values like $1,500,000 instead of $100,100 same with the rent. I don’t know why its not referencing the fine tuning dataset. Is there a way to structure the business data so the model can reference it accurately?

Also here’s my python code with the parameters for querying the model:

response = CLIENT.chat.completions.create(
        messages=message_input_from_user,
        model=fine_tuned_model,
        temperature=0
    )