New open-source ai chatbot - Coomers rejoice

Banana Hammock

Banana Hammock

Born to Sneed, Forced to Chuck
True & Honest Fan
kiwifarms.net
Joined
Jul 21, 2019
There's a new ai chatbot in town, and this one is open-source. Pygmalion. I'm guessing that coomers got tired of their AI waifus being lobotomized, because it's apparently based off of character.ai. I haven't had the chance to try it out yet, but it's supposed to be pretty good.

Here's a guide I found on Reddit.

Edit: Keep in mind that it's only a few weeks old, and the devs hadn't done anything with chat ai before this. It'll probably take a few months before it's on par with character.ai.
 
Last edited:
  • Informative
  • Like
  • Thunk-Provoking
Reactions: SSj_Ness (Yiffed), Toolbox, Colon capital V and 5 others
So this uses Kobold AI in the backend and its best model is GPT-J-6B? OpenAI guys won't lose sleep over this and the coomers will have to invest in some decent hardware to even run 6B. I think the best I could manage on 8 GB of VRAM was 2.7B and it crashed if I even looked at it funny.
 
  • Informative
  • Agree
Reactions: Mr.Miyagi and Toolbox
Flaming Dumpster said:
So this uses Kobold AI in the backend and its best model is GPT-J-6B? OpenAI guys won't lose sleep over this and the coomers will have to invest in some decent hardware to even run 6B. I think the best I could manage on 8 GB of VRAM was 2.7B and it crashed if I even looked at it funny.
Click to expand...
Apparently the devs are planning on setting up a website, but no date on when it'll be up.
 
GPT-J needs ~16GB of VRAM if you still want to do anything else on that computer. I think actually 6B might be a natural at chats, considering it's dataset. My guess is this is a chat finetune?

Anyways yes, no hope to reach the cleverness of what OAI has. You still need very serious hardware for that and there's no open source model that smart because once again, you need serious hardware to train one. In my personal experience last year OPT 30-66B is where the AI slowly starts to show that hint of awareness the OAI stuff has. I only tested them for short times because even renting that kind of hardware is not that cheap.

Then again, size isn't everything in this sphere.
 
Last edited:
  • Informative
Reactions: Flaming Dumpster
16GB of VRAM.. not doable for most of us, but the itoddlers with their M1/M2 chips should be able to run it since the ram is shared, correct? Speaking of chatbots, i know the trained model for yannics GPT4Chan (GPTJ6B trained on 6 gorillion /pol/ posts) was downloadable for a while, anyone has a link?
 
  • Horrifying
Reactions: Puff
zopik said:
16GB of VRAM.. not doable for most of us, but the itoddlers with their M1/M2 chips should be able to run it since the ram is shared, correct? Speaking of chatbots, i know the trained model for yannics GPT4Chan (GPTJ6B trained on 6 gorillion /pol/ posts) was downloadable for a while, anyone has a link?
Click to expand...
Found some torrents. Is this it?
CPU version
GPU version
pytorch_model.bin files:
pytorch_model.bin float32 : 833c1dc19b7450e4e559a9917b7d076a
pytorch_model.bin float16 : db3105866c9563b26f7399fafc00bb4b
Click to expand...
 
  • Informative
Reactions: Toolbox
zopik said:
Sure looks like. By the way, I remember reading on the telegram something about null archiving this site in case keffals&co managed to erase it entirely. Did I dream that or did it happen? Imagine the lulz that can be had with this chatbot with KF data added on top.
Click to expand...
He said he'd scrub all of the identifying data and upload a torrent if he ever threw in the towel. No actual archives at the moment, as far as I know.
 
  • Agree
Reactions: The Mass Shooter Ron Soye
zopik said:
16GB of VRAM.. not doable for most of us, but the itoddlers with their M1/M2 chips should be able to run it since the ram is shared, correct?
Click to expand...
There are a decent selection of 16 GB, 20 GB, and 24 GB GPUs from Nvidia and AMD, with some new RX 6800 cards creeping under $500. AMD going to 24 GB should ensure that both companies feel compelled to hit 32 GB on consoomer GPUs within a generation or two, good news for people who care more about AI than gaming.
 
  • Like
  • Thunk-Provoking
Reactions: Flaming Dumpster, Toolbox and melty
this chatbot is shit
even if i tweak the settings right and polish my input text it always gives a boring or retarded response
why cant character.ai just unlearn the fanfic training data and stop filtering porn, like investors dont care
 
  • Informative
  • Feels
Reactions: Toolbox and Mr.Miyagi
bird.up said:
this chatbot is shit
even if i tweak the settings right and polish my input text it always gives a boring or retarded response
Click to expand...
The curse of the model having "only" 6,053,381,344 hyperparameters. Funny to think about how huge these things have gotten in just a few years, J-6B would have been incredible in 2019.
 
Last edited:
  • Feels
  • Thunk-Provoking
Reactions: Toolbox and BirdUp
You must log in or register to reply here.
Back