Episode: 4386
Title: HPR4386: Silly Tavern Spicy Roll Play
Source: https://hub.hackerpublicradio.org/ccdn.php?filename=/eps/hpr4386/hpr4386.mp3
Transcribed: 2025-10-26 00:03:25

---

This is Hacker Public Radio Episode 4386 from Monday the 26th of May 2025.
Today's show is entitled, Silly Tavern Spicy Roll Play.
It is hosted by Operator and is about 20 minutes long.
It carries an explicit flag.
The summary is, Silly Tavern Spicy Roll Play for DeGenerates.
Hello and welcome to another episode of Hacker Public Radio with your host, Operator.
Today I'll be talking about Silly Tavern, more specifically, AI chat bot based.
What do they call it?
Roll Play?
The idea is that you get a large language model and then you build these character,
cards and then the JSON format and then you tell the AI to create a story and then there's various
controls and things you can do within the platform to figure out a kind of help guide you
in the direction to you want the story to go or whatever.
So I will actually run Silly Tavern, I haven't run it in a while.
Now part of jamboree jamboree.armacardi.com has a single click Silly Tavern experience
or whatever you want to call it.
So let me start it up here or download Silly Tavern or download jamboree.
There's a couple of them in the Kolbukai.
I started with stablehorde donating my GPU to hold on.
Donating my GPU cycles to image jamboree generation and then I figured out there's some roll
play aspects to it. I started playing when I was Silly Tavern and making spicy, whatever they call it,
spicy novels with my wife as a character in them and then found other repos of all kinds of
characters to as examples and you know there's all kinds of vial things on there that are kind of
crazy. So let's see let me start it up here. It takes just a few seconds to install.
So what you can do is you run my single click Silly Tavern.
I have pulls down job before you and get and pulls down whatever it runs it.
Now I'll start up the web UI. See if I can remember if it's I think it's still set up to talk
to my no I think I have to run the all talk so you can do a studio too. With all talk you can
have the audio get saved to a TTS fire provider fail to run so I don't I'm not using I don't
know where all talk is I guess it's on I said they're installed so the idea is you have these
character cards and they define by Jason the tag so it's got you have description which is
the description of the character then you have first message which is like their first matches
and I think you can have multiple ones and it's like a RNG so there's like three looks like this
one has three random first messages yeah had been doing had been will finish watching so it's
like basically scenarios essentially which is also confusing because each character can have
their own scenario I think yeah scenario of this character is characters husband which there's
a Jason element that's called open curly bracket curly bracket care clothes curly bracket
clothes curly bracket and it's like GAR so that's character and then there's oh with the other
one is user there's character and user user is you the person which I don't type I just have
characters talked to each other which can be confusing because character character so you have
to like replace I figured out the kind of the easiest way to do it is to replace character or user
with another character you have a name for so you basically have to have regenerate your character
cards every time you change what characters you're talking to are talking to themselves
so initially the setup for me has been number one the easiest thing to use is the built-in AI
Horde which is a free thing you can get kudos by donating your GPU cycles you can set that up
I can send you kudos I have a like I don't know two million kudos or something ridiculous if you
want some kudos reach out to me I can send you something to play with it makes it the generation
pretty pretty a little bit faster but it's not necessarily that much faster with text generation it
is with image generation so if you want to mess around with image generation or whatever
so I have it set up with local alama right now but ease of use you can use AI Horde then the next
thing you want to set up is well the context size I put the context size at 16 384 which is the
multiple of ten twenty four you probably want to start with whatever the default is eight eight one
nine two oh crap I lost my setting so now I have to do eight one nine two times two eight one nine
two times two because I use local models so I get 16 384 context token size response token size
at least small that's the response you don't want it to be huge and large and take forever to
give it to text the second part I went into characters but I'm trying to go into the basic setup
as far as I can tell now the other piece that I discovered has to do with at least the setup
that I had that's Apple H up whatever is that instruct template most of them follow that theme
but there's different ones like if you're using open AI it has different ones they have one for
gimme two they have one for vacuna and story I don't know what that pigman so these are different
instruct templates used for fine tuned AI models around roleplay I think where they have a
special basically different from different variables for different things so anyways the piece I
found is under system prompt the prompt content basically has some variables in there this say
something about user blah blah user user and character parameters which I never talk as user so
if it has user and character then I'm being thrown into the chat and I don't want to be I want
to be between two parts so the prompt prompt content field inside a system prompt under the
a settings advanced formatting I take the word user out basically anywhere I see the word user
I replace with a real character name in the in the whole setup wherever I find it and you can
see the debug stuff in a llama or you'll see it running in the background when you run silly
tavern which temporary let's see the other piece is a scenario I think scenario comes with the
character because this one says scenario husband is working abroad he feels lonely and scared alone
so like that I didn't type that in so it must have come from the characters
and then other setup is essentially the all talk for text generation that's pretty much all I
set up and that's a different conversation altogether I use all talk generally the other one I'll
use besides all talk is it that's pretty much that's it XCTS SV2 is is one I was using but I
use all talk for everything because it supports multiple engines so you can spin up multiple
doesn't matter what engine you use it all talk usually just works so and it does do the zero shot
cloning now cloning is you can fine-tune a model a pth or pt file which takes a fair amount of GPU
you need to have like a 16 gig card or more I think it might have all changed because the models
are getting tiny tiny tiny tiny and everything is getting very very very fast the model for
karaku karaku is like two gigs or something ridiculous and it sounds wonderful and it's
extremely fast because it's a small model so anyways um the all talk is a tts provider um yeah
I run just the bat script I think I don't think I need a jimpery for that it's all self-contained
using portable uh kanda mini kanda mini kanda and then there's a little smiley face that's called
personal persona management and I name my persona kiki q123 so that I can make sure that it
um to keep an eye out for it if it shows up in any of the chat and then you have your character
cards is where everything comes into place you have descriptions you have these tags which I don't
think are actually affect the character at all I think it's for like searching so if you wanted
to search for like female or not say for work or whatever um do you have it in there then you have
description and then you have first message like I said multiple messages and then there's all
this other weird stuff which I don't understand you have advanced um stuff that has scenario
and personality summary and character note and like example of dialogue which I don't understand
where any of that fits in I haven't messed with it enough then you have like scenes so like you
can get a world that's like you know uh you know Lord of the Rings and like fallout so you can get
a fallout template and play put like Lord of the Rings characters in a fallout scenario and make
like a cool thing and then change the characters a little bit to match like your own personality so
I have one from my wife in me and it's it's kind of how I'll make like cool interesting little audio
books using all talk um that's pretty much it and then there's lore books which I think are the
same thing can be world info um if you want characters pre whatever I went on GitHub and I'll see if
I can find the script am I still recording hopefully um I'll see if I can find the script to download
all the to pull down all the stuff it's time I get hub under a node so scripts node and it says
top 100 female and top 100 male and then I have chub AI character or dot org ripper dot js
it's a JavaScript that will rip all the
top characters from AI chub which is a bunch of examples of not great say for work uh you know
very uh not say for worked characters tall types of uh grotesque things that kind of pretty much
sums that up um now you know like some some of these will get very wordy
so you have to kind of guide it and there is something once you create a chat
I said it's been a while once you like start a chat there's something called our author
no authors notes and that's inside of like once you begin a chat there's thing called authors
you can format an author's note a default author's note it's applied to every single
new chat and the one I have is basically a jailbreak for roleplay and it's like assume the role
of a fictional character fictional roleplay always stays in character avoid repetition drive roleplay
by initiating actions do not talk about bonds the future do not talk about pushing boundaries
do not talk about making things unforgettable and experience in any way phrasing concept
above all focus mainly on responding to the user and it talks about graphic and explicit and
you know um uh vulgar and offenses and all that so like it's basically a jailbreak prompt for like
an old open AI or uh jailbreak prompt for something um and so you can put whatever you know grotesque
things uh horrible you know derogatory whatever I'm not condoning that but um it makes for some
some colorful uh some colorful chats and that can yeah where's the thing you can say
yeah in chat every five has system so like every four prompts it injects it back into kind of
guide and keep it on course so you know it can get repetitive or it can get um
kind of lost in the loop that's what happens with a lot of AI uh Andy does that in Minecraft um
auto GPT did it way back this did it too where it will get in a conversation with itself
um a lot of that has to do with like context size being too small or the uh not having the
authors notes or not having some prompting in there that will keep it from kind of repeating
so along with that you know um that's uh whatever and you know this can be considered
you know if if if I were to post like this jailbreak it would probably be against the
terms of of uh service for for HPR because it's got stuff like
vogal or offensive hateful explicit perverted derogatory disgusting and formal violent language
is allowed and can be used without when relevant and suitable for the character in scenario
so um you know explicit vogal graphic um for sex intimate body parts close
slight smell texture tastes sound to liquids feeling actions i don't yeah so anyways um it's like
what does it say how we tokens 348 tokens of just a jailbreak prompt for you know open AI
or something from 100 years ago so you can use jailbreak prompts for a roleplay to inject into
authors notes to keep it on track and not go bananas or in some kind of loop um but yeah I've made
it like a some like audio books to get to my wife and she was not impressed and didn't think
it was funny at all um but anyways uh that's pretty much it it's part of jamboree it's single click
now there's other ones now i'll post my AI notes let me make notes to make notes
oh crap before i forget uh so uh jamboree wow jamboree link with silly tavern button
then you want the link to a chub ripper then you want the link to other chats
now there's also my open AI notes which are just updated for a local
uh high school AI fix fabric roleplay there's an a g n a i dot chat
which is should still work yep and then there's light dot cobo kai AI dot net
so l-i-t-e dot k-o-b-o-l-d-a-i dot net it looks like that still works too and it uses by default
they both use open uh they both use open servers so you can click on horde here
trusted workers only or api key so i can say select models any and then it says i've got
it's 15 people no one count one so we can pick one that has like a lot of counts which means
that's how many servers are up with that model because you don't want to switch models it might
actually make something interesting to switch models around in the middle of your in the middle of
your thing so what if i if i did like select all models select models like select all of them
see what happens 29 oh wow that i don't know we can do that anyways so we've selected models and
like you got you i settings voice you can use local voice for all that subscription why
subscribe okay whatever um so you got you there's something name anyways i thought i could get
this going but it's look complicated robot and then like add a character you can use AI to build
your character it's like a whole thing i guess um so anyways um those are the two sites to play
around with it um and then silly tavern dot ai is like a uh they're they're kind of fan built
site for documentation on it but something to play on especially for your dnd person or if you
want to create like interesting stories you get these context windows for uh google that are
what what do they have to do two million or something ridiculous now so you can take some of these
characters in these methods and put them into whatever and basically generate a book
out of you know out of a few prompts and create a few characters and you know storylines and
kind of guide how the story goes um unsubrized learning how to episode about how to do that like
we're getting close to being able to create books with AI they're somewhat decent
anyways that's pretty much it for silly tavern um any other kind of chat based anything um
like side of use i use fast uh co fast cobo cobo lee whatever um i use fast fast
just fast co co co co co co co co co co co co co co co co co dash fast API for use with open web UI
because it has open AI based open AI compatible output english japanese korean chinese and
vietnamese coming soon um this stupid fast the model that uses is like i don't know what um like a
newer model um this wicked fast don't think you can use it with silly tavern the fastest one for
silly tavern is probably all talk with like fast tts or whatever it's called five tts v2
the voice cleaning is pretty good i've done manual training it's not that far off of the one
shot stuff but if you want to make a really good model with like inflection and stuff you need to have
you need to actually train it um that's pretty much it hope this uh has fun and you know
don't be too much of a degenerate right dig dig dig later
you have been listening to hacker public radio at hacker public radio does work today show
was contributed by a hbr listener like yourself if you ever thought of recording podcast
and click on our contribute link to find out how easy it really is hosting for hbr has been kindly
provided by an honest host.com the internet archive and rsings.net on the sadois status today show
is released on their creative commons attribution 4.0 international license