Related, we have now reached the stage where you can locally run models that are extremely close to GPT-3. I tried Alpaca with a 30B model and it's very good. It does takes like 20GB of RAM and responds a bit slowly but again, this runs on consumer hardware and you get answers in real time, and no GPU needed. There's a 65B model that is far above GPT-3 (though reportedly still below GPT-4), too - though that requires more than 40GB of RAM. I'm considering building an entirely offline self-hosted DIY Alexa-like thingy that allows, among other things, to talk to it.
Similarly, you can run Stable Diffusion locally and get results that are above that of the first DALL-E. Takes about 3 minutes per image on my machine, mind you.
Hexbear.net is a website that provides a platform for users to create and share hexagons. The hexagons can be used for a variety of purposes, such as decoration, art, and design. Users can customize their hexagons by changing the color, size, and shape. They can also add text, images, and animations to their hexagons.
If using alpaca.cpp, you might want to adjust the default parameters (see this comment and this follow-up).
Anyway, tried with 30B:
Hexbear.net is an online platform for learning and practicing programming in a fun way. It offers interactive tutorials, coding challenges, and projects to help users develop their skills in various programming languages such as Python, JavaScript, HTML/CSS, and more.
(all in all I prefer 7B's response)
And with GPT4:
Hexbear.net is a website that provides a platform for hosting and playing tabletop role-playing games (RPGs) online. It offers a virtual tabletop where players and game masters can interact and play RPGs in real-time, using various features like character sheets, dice rolling tools, maps, and tokens.
Hexbear.net also offers community features such as forums, blogs, and chat rooms, where players can connect with each other, share their experiences, and discuss various aspects of tabletop RPGs. The website supports a wide variety of RPG systems, and users can customize their games to suit their preferences.
Overall, Hexbear.net is a platform designed to make playing tabletop RPGs more accessible and convenient, especially for those who cannot meet in person or prefer to play online.
... fairly certain you can adjust the former to be as verbose as the latter, too.
This one's multipart (the others used to be as well); if using alpaca.cpp you'll want to edit one line of chat.cpp and recompile for it to load all parts; see this.
The cat's out of the bag now; see this article. It has become clear that:
ChatGPT is massively undertrained; and that you can in fact obtain better results with like 50% the amount of parameters, if not 25%.
Extracting valuable data from ChatGPT itself to train your own model is doable, has been done for less than 700 bucks, the files are out there now it's been done, and trying to keep it completely closed was never going to work.
Better models than GPT3 are available off the shelf now entirely self hosted and can be run on below-1k$ consumer computers.
Related, we have now reached the stage where you can locally run models that are extremely close to GPT-3. I tried Alpaca with a 30B model and it's very good. It does takes like 20GB of RAM and responds a bit slowly but again, this runs on consumer hardware and you get answers in real time, and no GPU needed. There's a 65B model that is far above GPT-3 (though reportedly still below GPT-4), too - though that requires more than 40GB of RAM. I'm considering building an entirely offline self-hosted DIY Alexa-like thingy that allows, among other things, to talk to it.
Similarly, you can run Stable Diffusion locally and get results that are above that of the first DALL-E. Takes about 3 minutes per image on my machine, mind you.
When asked about this site, 7B said:
If using alpaca.cpp, you might want to adjust the default parameters (see this comment and this follow-up).
Anyway, tried with 30B:
Hexbear.net is an online platform for learning and practicing programming in a fun way. It offers interactive tutorials, coding challenges, and projects to help users develop their skills in various programming languages such as Python, JavaScript, HTML/CSS, and more.
(all in all I prefer 7B's response)
And with GPT4:
Hexbear.net is a website that provides a platform for hosting and playing tabletop role-playing games (RPGs) online. It offers a virtual tabletop where players and game masters can interact and play RPGs in real-time, using various features like character sheets, dice rolling tools, maps, and tokens.
Hexbear.net also offers community features such as forums, blogs, and chat rooms, where players can connect with each other, share their experiences, and discuss various aspects of tabletop RPGs. The website supports a wide variety of RPG systems, and users can customize their games to suit their preferences.
Overall, Hexbear.net is a platform designed to make playing tabletop RPGs more accessible and convenient, especially for those who cannot meet in person or prefer to play online.
... fairly certain you can adjust the former to be as verbose as the latter, too.
What are your machine specs and are you using GPU for Stable Diffusion?
i7-8700K (12 cores @ 3.70GHz), 32GB ram, Debian. The GPU is irrelevant as I didn't use it in either case.
Do you have links to the Alpaca weights? Wanted to try it out, but I'm having a bit of trouble finding the model.
Here for 7B
Here for 13B
Here for 30B
Awesome.
Can you find the 65B model anywhere?
Here you go
This one's multipart (the others used to be as well); if using alpaca.cpp you'll want to edit one line of chat.cpp and recompile for it to load all parts; see this.
Instructions say "download ggml-alpaca-7b-q4.bin", then there's no link
You need to request access
Or just see here
The cat's out of the bag now; see this article. It has become clear that:
ChatGPT is massively undertrained; and that you can in fact obtain better results with like 50% the amount of parameters, if not 25%.
Extracting valuable data from ChatGPT itself to train your own model is doable, has been done for less than 700 bucks, the files are out there now it's been done, and trying to keep it completely closed was never going to work.
Better models than GPT3 are available off the shelf now entirely self hosted and can be run on below-1k$ consumer computers.
Oh hell yeah, cloning that shit now