Train Your Own Private ChatGPT Model for the Cost of a Starbucks Coffee | Hacker News Reader

Train Your Own Private ChatGPT Model for the Cost of a Starbucks Coffee

medium.com

73 pointsDSOfficial17 comments

This isn't anywhere near equivalent to ChatGPT. I am immediately distrustful of any blog with a title like that.

To my knowledge, it isn't possible to self-train anything anywhere near equivalent to ChatGPT GPT-4 model, right? What they have is proprietary and subsidized by Microsoft letting them have access to probably... hundreds of millions of dollars in free Azure cloud compute/bandwidth credits + customized hardware?

With the current technology and methods, correct, you have to have tons of money that you would rather gamble than invest.

It kinda glosses over the part that interests me the most - how one would go about doing this part:

>Domain-Specific Knowledge:

>For organizations or individuals with specific business requirements, training your own ChatGPT model ensures that the model has the latest and most relevant knowledge related to your business. Regardless of your business domain, a model specifically trained for your business needs will be more valuable than a generic model.

Say I want to train it to filter news articles for me. How would I go about building this dataset of examples of what I want to see?

browser history, perhaps? or upvoted post titles from HN?

Not just the source, but like what do I do with it. Say I've got my list of urls of "good" news.

I'm guessing I need to somehow download the content, standardize it in some unspecified manner and get it into some unspecified format before putting it into this folder they're talking about there.

Very much feels like a "rest of the fuckin owl" situation of "train your own" model. It's missing the key part of "own".

I can't speak for the technique in the post, but you could possibly use https://cohere.com/ to do the sort of classification job you mention. The training appears to be free (at least, mine have been) but you pay for the eventual classifications. It's an easy process to do the training and play about though. If you meant something more elaborate by "filter" than "classify" then ignore this comment :-D

With the cost of a cup of Starbucks and two hours of your time, you can own your own trained open-source large-scale model. The model can be fine-tuned according to different training data directions to enhance various skills, such as medical, programming, stock trading, and love advice, making your large-scale model more “understanding” of you. Let’s try training an open-source large-scale model empowered by the open-source DolphinScheduler!

It will even write your peppy upbeat promotional posts for you!!!

I'm amused how "chatgpt" is becoming genericized.

so now when your buddy approaches you to build his next big app idea, it'll be "chatgpt for hotdog toppings" instead of "uber for hotdog toppings"

tldr: LoRA fine-tuning of an open-source alpaca model.

See https://arxiv.org/abs/2305.15717 for why calling this a private ChatGPT is probably misleading.

(Based on alpaca-lora)

> for the Cost of a Starbucks Coffee

So like $400

is it self-host-able?

TL;DR: Cheap GPU cloud with alpaca-LoRA

How can I deploy it?

Reply on news.ycombinator.com