Pygmalion AI

[ThexOrion]

I know that a number of people have talked about this on the character.ai message board. I wanted to make one that was just for Pygmalion.

If you don't know, Pygmalion Ai is a web based ai chat program in a similar vein to character.ai. The difference is that even though character.ai works for the most part, it has a nsfw filter that is pushing people towards Pygmalion, which does not have an nsfw filter.

People can create their own characters in Pygmalion, but these are stored in the form of .json files instead of being stored on the website like with character.ai. I wanted to create a board for people to talk about Pygmalion, as well as post links to characters.

Link to the website that has both some info and a link to the google collab that can run the ai.
https://rentry.org/pygmalion-ai#play-wi ... odels-here

Warning that this does require some basic technical knowhow to use properly.

[Typher]

you can set it up locally ? that's a big yes from me, i'm going to try and set it up, i alway wanted someting like character.ai that i could run locally

[Bigfatpiggy]

Ah yes, pig mailman AI.

[grandplex]

Whoa there. This is mostly wrong. Pygmalion isn't web-based at all, though that is planned eventually, and it can be setup locally, but only with a monster of a computer along with a lot of knowledge, or by using Tavern AI to save your chats locally, along with Colab or Kaggle+Kobold.

The one that gives you a permanent web like interface is when you combine Pygmalion+Kaggle+Kobold+Tavern (and some other support programs), and the outputs can be pretty incredible, such as page long detailed responses to some simple prompts. It is free, and won't melt into one-liners or gibberish like characterAI does if you talk too long to it. That all said, on a scale of 1 to 10, I'm probably around a 8 or 9 with computer knowledge with a moderately cutting-edge computer, and it took me a couple days to figure out the right setup. The setup instructions are on reddit.

[dst3313]

You can set it up locally, but you need a beefy computer. It requires a minimum of 7 gigs of video ram I believe. I know this because I have 6 gigs.

[grandplex]

The latest 6B recommends 16GB VRAM, I believe I remember it using 12GB using Collab+Kobold. That means you need the 12GB version of a 3080 or better, though that is very borderline, the recommended is a 3090/4080 equivalent or better, just one small step down from the fastest cards you can buy.

[SputnikDX]

Also to the characters don't "learn" similar to the way Character AI does. When you start a new chat with Character AI, the AI has learned somewhat from the previous chats. However, Pygmalion does not do this, and every new chat is a completely blank slate from what I've seen.

[Halowinner]

I tried it but unsure how to get it to do vore or even other vore type also sorry for necroposting i guess?

[grandplex]

It can do vore as long as you put it in the character description, that said, for one thing, it likes to almost always screw around with wordplay a lot before actually doing it unless you really force it to start with your own prompted words, and it also is still a somewhat less descriptive than CAI (and I don't think CAI is great at it either, it is just slightly more descriptive and is much more willing to immediately do it, on the other hand, the filter lobotimizes it so much that it can't describe half of what it should so CAI's potential descriptive ceiling is really low, painful to push against, and will probably get worse).

[DragOnDude]

The colab book worked so well at the beginning and now it won't be loading my json...

[Trubot527]

The colab book worked so well at the beginning and now it won't be loading my json...

There are usage limits, so you can't just keep it running all the time. It refreshes after like 24 hours, but when you run out of the usage time, it relies on your own computer to function, which would be extremely slow, unless you have a high-tier computer. You could use alt google accounts to bypass the limit though

[grandplex]

About 3-4 google accounts is enough for huge RP binges each day.

In other news today, for Pyg and other models, there was some mention of them simplifying the power requirements for the really big models past 6B so they can work on what is available on things like collab, that said, it will take several weeks or months depending how big a training library they build for one of those higher versions. I would expect there to be some decent vore training eventually in one of those, so it can properly include vore actions, rather than just suddenly putting you in there stomach, or giving you a bunch of vore dialogue without any/almost no action.

[AlluringPredation]

If you mean FlexGen that's currently only for raw OPT models. So no additional fine-tuning is foreseeable. Also what that article didn't mention that their speed test was based upon a 512 token context where standard usage is a 2k token context which results in replies from the AI on their system to take 8 minutes for a 50 token output. Also required 208 GB system RAM which most people don't have available.

You could also mean the 4bit INT compression while greatly reduces GPU RAM usage it also severely slows down inference and somewhat reduces accuracy.

We're currently waiting to see if the pythia model pans out, or waiting until PileV2 is released and RWKV 24b gets trained (and hoping that BlinkDL decides to reincorporate the literary corpus from PileV1.)

Regardless, it costs thousands or tens of thousands to finetune these models, which also requires a nifty dataset.

[Panzershark]

Is Pygmalion still planning to have to website up soon or has that become a wash?

[Ottots]

Is Pygmalion still planning to have to website up soon or has that become a wash?

They said it was being worked on, but they never said it would be done soon. It will be quite some time before they get a site up due to the nature of ai technology.

[AlluringPredation]

Just wait for the Llama finetune

[bonberjean]

I used pyg with kaggle once (before kaggle nuked something that made it unsupported), it was very good for regular nsfw, so-so for furry nsfw and very bad for vore of any kind.
The AI loses context very quickly it seems.

[Newtonne]

Just wait for the Llama finetune

^This

Unless someone trains a model on all the vore stories on Eka's Portal (which would probably get a lot of people upset), we're just going to have to wait for powerful RLHF models.

[AlluringPredation]

Just wait for the Llama finetune

^This

Unless someone trains a model on all the vore stories on Eka's Portal (which would probably get a lot of people upset), we're just going to have to wait for powerful RLHF models.

It could be done, but only opt-in to avoid getting people understably riled. With a fine-tune the AI could potentially learn about the included characters and who they are, unlike the module discussion in the past where it just learned language style.

Do note the major problem being it is against the license agree of Llama to share the model with others, it is "supposed" to be only available through them directly..

[Molbo]

Room & Proxy: https://github.com/Cohee1207/SillyTavern / https://alwaysfindtheway.github.io

Chara: https://www.characterhub.org/characters / https://booru.plus/+pygmalion / https://mega.nz/folder/FWgxyaLA#1Dxz2IYVibPBSMnM1Rut7Q / https://discord.com/invite/pygmalionai