Fine-Tune Data, [text] , [label] not allowed

i have chat with ChatGPT and has it teach me how to convert certain data into OpenAI GPT-3 fine tune data format in jsonl.

it show me some example included those look like this

{“text” : “ABC is a …” , “label” : null }
or “label” : “company description”

not the typical Prompt : , Completion
fine-tuning data.

but when I submit to OpenAI via API,
error message show me
the 1st line of data missing Prompt, and completion.

Is ChatGPT wrong about this?
all fine tuning data via API Must be

Prompt + completion?

is Prompt + Completion must always be in Q & A format?
Prompt = Question
Completion = Answer?

Is there other example?

Thanks in advance.

Hi @zhihong0321

I have cautioned you kindly a number of times to not use ChatGPT for these types of technical details.

Please refrain from using ChatGPT to write code for you and read the API docs and learn to code from the examples in the API docs.

ChatGPT is not “teaching” you anything, but most wrong things because ChatGPT is a language model, generating text, blah blah, not an expert system.

Do you understand this?

I ask you because you keep going to ChatGPT to get wrong answers and then you ask us to correct the wrong ChatGPT blah, blah based on the fact that we read the API docs and we “experienced developers” do not use ChatGPT to code.

What are are essentially doing is going to a kind of “idiot” to generate blah, blah, blah code which is wrong and then you come to us to “fix” the code you got from an “idiot”.

Why not just read the API docs and code using real, working examples?


Hi Ruby, thanks again your prompt reply.

i did know ChatGPT coding is not as accurate as it should be…
but… currently, for a nocoder like me, ChatGPT did enabled me for many possibilities.
I just need chatgpt give me quick idea for any question. and try and error… viola… magic happen.

A great example :
I was building a POS system on , which previously rely on a quite sucked paid plugin but no choice ( ESC/POS printing from Bubble directly )

i try my luck to have ChatGPT write me a javascript that can have my site to send ESC/POS command to connected printer directly. i have new progress.

ChatGPT may not prove useful for experience Coder…
But… it just enabled non-coder 1 big step to more coding possibilities.

---- back to your reply

I have read the API Doc many times.
It only provide example of Prompt , Completion combo

I was thinking if i going to update long description content…
do i need to split it into multiple Q&A ( prompt, completion )?
or i can do it 1 shot?

This is not what I said.

I said that ChatGPT generates a lot of bad code. It makes things up. It hallucinates.

The difference between me, an experienced coders with decades of coding experience, and you, is that I go to the API docs and verify everything.

In other words, I don’t create code using ChatGPT and then when it does not work, go ask others to fix it, unless I have exhausted all other options. This normally means reading some documetation.

That is because there is only, prompt-completion keys when you fine-tune.

There are no text-label keys like as follows from your post:

You went to ChatGPT and it generated wrong information. Then you went to the API does and of course it does not have the information from a totally wrong ChatGPT completion, because what you are working off of wrong initial ChatGPT info. That, my friend, is not “teaching”. That is misinformation.

Honestly, @zhihong0321, I don’t see where you are more "enabled’ by ChatGPT as a novice coder. If you go to the API docs and follow the examples in the OpenAI API docs you will be learning the correct ideas and you will learn to code OpenAI API tasks faster, not slower.

When you go to ChatGPT, you get a lot of. blah, blah, blah like this:

The above is wrong. ChatGPT pre-trained data ended in mid 2021. This is 2023 and a lot has changed. You are working off old, depreciated, and often totally hallucinated ChatGPT completions.

OK. Take care @zhihong0321

I now understand you are a copy-and-paste ChatGPT coder and you are not going to follow me as I have advised you many times and go study the API docs and follow the actual working examples. You are going to keep replying on ChatGPT and this is going to slow you down considerably.

Please understand. We are busy people. If you insist on getting your code from a hallucinating chatbot with outdated information and deprecated code, and then ask us to correct all the mistakes that ChatGPT makes, it’s not the right approach to learn as a developer, even as a novice.

You should work the current examples based on the current code and current API methods on the current and up-to-date OpenAI platform site.

Learn the correct way, is what I advise you. Save “chatgpt coding” for later when you master the basics.

Best of luck!


Thanks for providing the Answer. :slight_smile:
When ChatGPT cant help. You are my last line of Hope :slight_smile:

I have another question.
regarding Insert, and Edit.

could you explain it further?
especially in the case I needed to “customize” the bot.

What is {company name}
{company profile goes here}

I have try fine tuning. (curie, 16 epoch)
i prompt the same question, but answer i get :
., a
HiOrder system.###… you to the, for, and,Order

Original Finetune data:
{“prompt”:“what is hiorder?\n\n###\n\n” , “completion”:" HiOrder is a Smart Restaurant Cloud System contain 3 module, which is QR Menu for ordering automation, a POS (point of sales) and a Kitchen Display System (KDS). It allows restaurants to automate and operate more efficiently by taking advantage of the cloud. \n\n###"}

does Edit the faster way to achieve this?
Or this is not what Edit does.

Thanks, but I’m not going to continue to be your “correct the hallucinating chatbot” on-line buddy :slight_smile:

No. Forget about the edit endpoint for now :slight_smile: I have never used it, personally.