h2oGPT - The world's best open source GPT
-
h2oGPT - The world's best open source GPT
https://github.com/h2oai/h2ogpt
Our goal is to make the world's best open source GPT!
Try h2oGPT now
Live hosted instances:
For questions, discussing, or just hanging out, come and join our Discord!
Current state
- Open-source repository with fully permissive, commercially usable code, data and models
- Code for preparing large open-source datasets as instruction datasets for fine-tuning of large language models (LLMs), including prompt engineering
- Code for fine-tuning large language models (currently up to 20B parameters) on commodity hardware and enterprise GPU servers (single or multi node)
- Code for enabling LoRA (low-rank approximation) and 8-bit quantization for memory-efficient fine-tuning and generation.
- Code to run a chatbot on a GPU server, with shareable end-point with Python client API
- Code to evaluate and compare the performance of fine-tuned LLMs
All open-source datasets and models are posted on
H2O.ai's Hugging Face page. -
Would be great.
I don't understand if the nvidia bits are possible in a Cloudron image. -
@robi I've been playing around generally, and certainly hope this opensource offering becomes possible on Cloudron.
Just as feedback for anyone else here exploring AI, my main use of AI tools so far has been to help generate code for apps.
I tried the hosted h2oGPT but I have to say that I have not yet found anything better than ChatGPT4 through their paid $20/monthchat.openai.com
I am certainly hoping to find something, especially if it can be opensource.
So any suggestions / pointers are very welcome.
But for me, the hands-down winner for code generation at the moment is ChatGPT4.It's frankly beyond amazing.
At times, it's like having a senior developer acting as a tutor or mentor.
Most of the time, it takes me back to my days running software projects and writing requirements and functional specifications for developers to produce the code.
But ChatGPT4 is few hundreds times less buggy, a few thousand times faster, and a few million times cheaper than a human developer.
And you don't have to buy it coffee or doughnuts, or give it holidays or maternity/paternity leave.
[well, ChatGPT4 is rate-limited but so far not a problem, as it takes time to compile and test code, and work more on next stage of the functional specification]Lots of public discussion on AI highlights the likely impact on jobs, but the discussions I've seen are mostly about jobs in industry and general business.
I'm starting to think one of the hardest hit sectors is going to be developers.
At least developers in big dev houses.
Maybe there will be flip side that there will be a boost for independent developers and tech entrepreneurs.Sorry, I'm going off topic.
I'm very interested to hear about how others are finding different AI models and platforms for coding, as well as generally.h2oGPT has not so far worked well for me on this.
Although - fair play - it was a lot better than Google's pathetic Bard.
If anyone has found something better than ChatGPT4, I definitely want to know about it !
Maybe h2oGPT needs a different technique for prompts.If the Cloudron community can properly train or fine-tune a model on Cloudron packaging methodology, the much-discussed packaging bottle-neck might be solvable.
-
@timconsidine It's not h2oGPT that's not good at code, it's the model you selected combined with the prompts you use to elicit a successful response. These can vary from model to model.
You can upload the StarCoder model and see how it does, even use GPT4 to evaluate it.
There are even better ones that take StarCoder and fine tune it with OpenAssistant: https://huggingface.co/spaces/HuggingFaceH4/starchat-playground
Folks on YT & Twitter are often doing benchmarks and comparisons for all these.
-
And not a hint of embarrassment ...
h2oGPT 20B