Business account error while fine-tuning a model

Hi,
one of my colleagues subscribed a business account for OpenAI API, paid some money, created a Team and invited me as a member of that Team.
I joined the workspace and tried to fine-tune a model (gpt-3.5-turbo-0125) but when I click on “create” this error appears:

Error creating job: Fine-tuning jobs cannot be created on an Explore plan. You can upgrade to a paid plan on your billing page: https://platform.openai.com/account/billing/overview

Is it because I result as “member” and not “owner”?
Should I pay something as well?
Thank you, regards.

Roberto

ChatGPT (including the Team subscription) is a completely separate product from the API.

You’ll need to create an API account and add funds if you’re wanting to fine-tune any models, but you should know that any fine-tuned models you make will not be available through ChatGPT.

Before doing any fine-tuning, you should read through the documentation and make sure that creating a fine-tuned model actually fits your use case (e.g, fine-tuning does not add any new knowledge to models!).

Thanks for the quick reply.
As far as I know we paid for OpenAI API as you can see in the screenshot:

image

isn’t it correct?

Regarding the “fine-tuning”…I should “train” a GPT with my own data to generate a JSON structure: isn’t it made through “fine-tuning”?
If not, what’s the way to do that?
Thank you,

Roberto

You’re correct, that’s the API. You mentioned “Team” so I thought you might be referring to the team plan.

When you look at the usage page what Tier does it say you’re in at the bottom? It should look something like this:

As for the “fine-tuning” it really depends on what your use-case is. Why are you wanting to fine-tune the model/what are you using it for?

1 Like

It results “in free”

image

My “use-case” is a bit articulated…we are trying to integrate OpenAI with our program where a user has to compile a form with some fields.
The “schema” of that form is customized by the user so it can have several forms with different fields.
At the moment the user compiles the form manually through the keyboard but we would like to let the user “dictate” the fields values through a microphone.
For that we are using “MS Cognitive Services” that transform the spoken language into a text that we would like to “send” to the AI so that it generates a simple JSON structure.
For example…I have 3 fields “brand, color, size”…the user could say “the brand is brand1 with color red and size medium”…the AI should reply a JSON structure with the three fields and values.
The JSON structure will always be fixed, just a “list” of couple “field name:value”.
So my thought was to “train” the AI with the list of all the available fields so that it can recognize them exactly with their “name/code” (the name/code is an alphanumeric identifier).
How can it be done?

Rate limits - OpenAI API

To fine-tune models I believe you’ll need to be in at least tier 1, so you’ll need purchase at least $5 in credits (yes, unfortunately the $100 pre-loaded don’t count towards this).

To be honest you’re always better off trying to do it using normal prompt-engineering in my opinion. Inference on fine-tuned gpt-3.5 is six times more expensive than the base model, so it may be more affordable to just write a few-shot system prompt.

My recommendation is to just try creating it using the base model and refine your prompt to it (explain the use-case, provide it a list of the different acceptable fields, etc.) and only fine-tune if the results are inconsistent or the prompts become very costly.

Hope this helps!

The API has its own organization member structure. Being on teams alone doesn’t affect that, you are still looking at your individual account’s API balance when you see “free tier”.

You’ll go to platform.openai.com, and see the sidebar with “organization” and “members” categories. The account that purchased API credits will be the only organization that can use those credits or fine-tune AI models.

However, that account can also invite other members to the API organization as “reader”, and then that member can bill their API usage to the default organization under API keys, or an organization can be specified when making API calls.

This will let you work with an organization’s fine-tune models and storage.

2 Likes

If I got it right, the one that can fine-tune through the “API web site” is the owner of the organization, invited members can’t.
But if I “fine-tune” a model by a call to the API through my program (providing the organization’s API keys) I could do it.
Is that right?

Roberto

I’ve read the articles about fine-tuning but…I’m still a bit confused.
If I got it right, fine-tuning is used to “refine” the response of the AI so that it can be “more specific” about the topic of the user request.
To do that we need to supply a JSONL (or other source) that contains all the possible requests/prompts with their respective response that will be used by the AI.
So…how to “teach” the AI to do something new?
This point is confusing me a bit…

The application people typically want is “answer concretely about my products or price list”, or more nonsensically “chat with my PDF”.

That is best done a search or vector database lookup based on the user input against knowledge documents, then placed into the AI messaging input before an answer is continued; not by the tedious process of teaching an AI how to produce and infer response sequences in a different manner.

1 Like

I’m experimenting with “prompt engineering” as suggested by an user.
I found a prompt with which I can obtain the response I expect, I just found a strange issue: if I ask the AI “what’s the date of today” it replies with a wrong date, like, for example, “2022-01-21”.
Is that normal?

Roberto

Edit:
I found the answer myself to the date problem…the AI can’t reply with the current date through the API…a workaround is to include a system message like, for example, “today is 2024-03-27”. With it AI can even calculate past or future date.
That “date issue” is not present with ChatGPT.