ChatGPT fine-tuning as a service

Hey all,

What do you think the likelihood is that OpenAI will offer a service to fine-tune chatGPT on your company’s own data? Feeding it your company’s code, documents, slack messages, etc… will largely solve the documentation problem that most businesses have. AWS and Google already offer services like this, but obviously they’re not as good as chatGPT yet.

1 Like

Go watch Sam Altman’s keynote from a few weeks ago. He says that he expects that the next unicorn companies will focus on fine-tuning models for various verticals. I think OpenAI is going to focus on building these LLM, and let third-parties deal with the fine-tuning.

I’m exploring gpt-index right now to build indexes of our company source code. So far I’ve been able to build an app that works on the command line and builds a working context index from my console input. I can load files into index, or directories, then query directly against it. It’s all prompt engineering, but I’m going to be incorporating results and refinements into the workflow so that I can curate and fine-tune certain models.

1 Like

@dahifi where is the keynote link, can’t find it.



What Sam Altman was referring to here is that people were going to take open source models and fine tune them (or re-train them) to do their own thing. This is already happening! I don’t think OpenAI is going to let third parties fine-tune their models, because I don’t think OpenAI will open source their models and let others run a fine tuned model without their API. It would be cool if they did. But we’ll see. I know they have open sourced some of their models, but their big models (GPT-3), no way.


i was wondering if there are any simple steps to fine-tune gpt-3 with a chat template that looks more like the one openai is using.
any ideas …

You can do this to a limited extent. ChatGPT for instance maintains a context vector of about 3,000 words to keep track of the current chat. That of course is not a lot but, it can be helpful to feed in new data (just ask ChatGPT to read it). I have been able to get it to write an article this way about a subject it did not have enough specific data for.


Yes, it will be just a time problem. Indeed, OpenAi already has their api for fine-tuning GPT-3 model, it has no reason they don’t release the fine-tuning api for ChatGPT.

OpenAI will be releasing an API for ChatGPT soon. It’s been announced by both OpenAI and Microsoft’s CEO Satay Nadella this week.

My guess is that the ChatGPT API will permit fine tuning.


If there will be API, the ChatGPT will be paid like GPT-3 is now?

Does anyone aware what’s the block size of chatgpt ?

Hi @jon.flynn2

I am actually working on that at the moment. I created a way for users to upload documents or add website to your own custom library and run GPT on top of it. Let me know if you like to test it out!

Let me know how, because I would like to try.

Sure~ Fine-Tune will provide more ML infra. incoming for OpenAI and Azure. You need to pay for the huge ML cluster (they built with lots A100 GPU) usage…

I think OpenAI and Azure will provide Fine-Tune model but may be not let you export it from Cloud platform… You can import data and re-train the models.

Hey guys,

i tried to train a few models, with best practise training approach. My expectation was, that a fine-tuned model extends the exiting models from openAI, so that i can use the abilitiy to communicate very well PLUS the knowledge of my fine-tuned data. Maybe its a silly question, but testing my fine tuned models results in nothing really great, in best way they find the source of knowledge related to the prompt and put something around that out - but not in high quality manner i would expext from GTP3-4.

Is it right, that a fine tuned model is NOT an extension of the whole openAI knowledge of the world?

What are you using to fine-tune? OpenAI has not yet released fine-tuning for GPT3.5/4 but they are planning to: OpenAI Platform

" they find the source of knowledge related to the prompt and put something around that out ". That sounds like a hacky workaround that everyone is doing to try to extend the model’s knowledge with a vector database. google Pinecone, Langchain, etc… to learn more about it. The idea is to index your data in a database, search it against your query and throw it all into GPT4 to summarise it. This is very different from actually changing the model’s weights and making it more intelligent, which is what fine-tuning means.