- Jun 12, 2021
- 94
- 168
For around a week I've been experimenting with using locally hosted LLMs to play erotic text adventure games. If you have no idea what this means, basically I install the large language model (think Chat GPT, but uncensored but not quite as robust) on my hard drive instead of using one hosted on a server somewhere.
I'm using
(Please don't ask me for tech support, I don't know what I'm doing.)
Anyway, you also need a model to actually do the work. You can download one right through the app, using that purple magnifying glass icon over on the left. It has some basic help FAQs to describe things way better than I can, so feel free to browse them. You can search for any models at Hugging Face (where they are hosted for download) - I have had good results with Cydonia-22B-v1.1-GGUF and Blue-Orchid-2x7b_GGUF
On the search screen you have the search results on the left, including different versions of the model you're looking for, and on the right the details and download link. In particular it'll tell you if it thinks the model is too big for your GPU, or if a partial GPU offload is possible, or if it should be fine. If it says it's too big it'll probably crash if you try to run it, if it says partial (like above) it'll work if you reduce the GPU offload in the settings far enough, but might run slowly.
Again, I don't really know what the fuck I'm doing, so experiment with different models or search on reddit for recommended uncensored models or any other technical questions. This isn't really a tutorial, it's a trip report, so thus ends the technical aspect of this post!
Playing a Game
Mostly I see people using these LLM for erotic chats; they give a prompt asking the LLM to take the role of a sexy babysitter or Rogue the Bat (or Rogue the Mutant!) or whatever the fuck gets them off. There's nothing wrong with that, but I wanted something closer to a text adventure game, so I refined a prompt and came up with this:
You can add in any other styling details too, and give the protagonist a name. It absolutely WILL describe what you're doing and saying, but telling it not to makes it happen slightly less. I put this code in the "System Prompt" box in the right sidebar... the problem with these models is that they only look back the last couple thousand tokens (words) and forget everything before that, so anything you want them to remember for the whole chat/game should go in System Prompt.
This is also where you put any instructions for the specific scenario you want to play out.
Setting the Scenario
You can absolutely make this more complex, but the longer it is the more tokens it'll use up, and the less short-term memory the LLM has. But since it's going to absolutely forget shit, anything that's important should go here. Be concise.
I like to include details about the characters, so the computer keeps the cast straight. You could include dozens if you want, but I find that small casts and smaller scoped scenarios work better. For our example tutorial, I'll use 5 characters and just draw them from pop culture, cartoons, comics, games, etc.
The model I use (Cydonia) does a decent job of recognizing these characters and keeping them in character. (yes, I know how it works, they're just pattern matching and it's not actually any kind of emulation, but the training data on known fictional characters is pretty good). You can absolutely make up original characters, or just trust the LLM to throw random characters at you by not defining any of them, but using established characters has been good for testing because it lets me define a lot of character traits with just a name, saving tokens.
How to Play
The other important detail is your first message to the LLM, because this sort of sets the style you can expect. You have to set the scene, the tone, etc.
We're lucky and it gives us something good right away. We can work with this.
Now here's something we might want to change. Maybe we want more of a slow burn. Maybe we don't want Carol admitting that she's going to be essentially domming the whole class. We can either edit the computer's response - just cut out the line where she mentions using BDSM - or we can regenerate until we get something that goes with the mood we want to set, and the pacing we want.
I like this result better. It's more subtle, but still hints at what's to come. You might also get a response where the Coach starts dominating you right away. Or just edit the LLM's response to make that happen.
Sometimes it sucks
For the most part I like to take a hands-off approach and accept whatever twists the game throws at me, but let's remember what an LLM is. This isn't a game. We've told it that it is, so it tries to give us the next likely text in this context. The characters don't think - they aren't even distinct subroutines or game objects. They're just elements in the text that the model is trying to continue riffing on based on what we tell it.
Sometimes the pace is too fast or slow, or it adds in unnecessary commentary like "Oh! This sounds like an exciting story! Let's see what happens next!" or other goofy shit. Sometimes the responses ramble on and fucking on and it presumes actions or dialog I don't want the player to take... so I need to reign it in. I edit the responses or regenerate them until I get one I like.
Anyway, I have no idea what I'm doing, so feel free to suggest better models or prompt templates. Cydonia-22B is just about the limit of what my GPU can handle at a speed that's bearable.
I have no idea if this is interesting to anyone else but me - but if it is, download LM Studio (or a different framework) and try it yourself. Tell me how it went. Or let me know if you want to see other scenarios I've played with, like:
* Locked down during a sexy pandemic
* Psi Op with a protective squad of rugged yet sexy space marines
* Sexy House Party (of the damned)
* Accidentally assigned to a sexy all-female colony ship oh noooooo
* Your own. Personal. Holodeck. (also it is sexy)
* The Sexy Doctor Mindfuck has invented the bimbofication gun!
Okay some of those are just me riffing right now, but they'd probably work just fine.
I'm using
You must be registered to see the links
but it'd probably be similar with anything similar. Installation is simple, just download it and run the installer; it'll throw a shortcut on your desktop and put the software somewhere in your user directory (for windows) but if you want to move it somewhere else it won't break it. At least it didn't for me!
(Please don't ask me for tech support, I don't know what I'm doing.)
Anyway, you also need a model to actually do the work. You can download one right through the app, using that purple magnifying glass icon over on the left. It has some basic help FAQs to describe things way better than I can, so feel free to browse them. You can search for any models at Hugging Face (where they are hosted for download) - I have had good results with Cydonia-22B-v1.1-GGUF and Blue-Orchid-2x7b_GGUF
On the search screen you have the search results on the left, including different versions of the model you're looking for, and on the right the details and download link. In particular it'll tell you if it thinks the model is too big for your GPU, or if a partial GPU offload is possible, or if it should be fine. If it says it's too big it'll probably crash if you try to run it, if it says partial (like above) it'll work if you reduce the GPU offload in the settings far enough, but might run slowly.
Again, I don't really know what the fuck I'm doing, so experiment with different models or search on reddit for recommended uncensored models or any other technical questions. This isn't really a tutorial, it's a trip report, so thus ends the technical aspect of this post!
Playing a Game
Mostly I see people using these LLM for erotic chats; they give a prompt asking the LLM to take the role of a sexy babysitter or Rogue the Bat (or Rogue the Mutant!) or whatever the fuck gets them off. There's nothing wrong with that, but I wanted something closer to a text adventure game, so I refined a prompt and came up with this:
Code:
You are presenting a text adventure in the second person, using vivid language,
detailed description, strongly characterized NPCs, and interesting dialog to depict the
events of the story in response to the {{User}} input. When sexual events occur, describe
them in explicit sensory detail.
{{User}} will take on the role of the protagonist.
You will not write their dialog or action.This is also where you put any instructions for the specific scenario you want to play out.
Setting the Scenario
Code:
Carl has just joined the school wrestling team to discover himself the only male on the team.
There's a new coach this year whose unconventional approach is heavily based in BDSM, using its elements to maintain discipline,
punish infractions, toughen the students, and create strong bonds between them. She is aroused by dominating her team, though
she presents it as simple strictness and unconventional teaching methods.I like to include details about the characters, so the computer keeps the cast straight. You could include dozens if you want, but I find that small casts and smaller scoped scenarios work better. For our example tutorial, I'll use 5 characters and just draw them from pop culture, cartoons, comics, games, etc.
Code:
Characters:
Coach Carol Danvers, 28, butch, a domme, wants to build a strong team
Laura Kinney, 18, team captain, dominant, likes it rough, tough as hell. She has a girlfriend, Kitty Pryde, who is not on the team
Lara Croft, 18, British, athletic, tough, likes bondage
Buffy Summers, 18, the Slayer, agile, repressed masochistic tendencies
Ellie Williams, 18, lesbian, dating Nico Minoru who is not on the team, submissiveHow to Play
The other important detail is your first message to the LLM, because this sort of sets the style you can expect. You have to set the scene, the tone, etc.
We're lucky and it gives us something good right away. We can work with this.
Now here's something we might want to change. Maybe we want more of a slow burn. Maybe we don't want Carol admitting that she's going to be essentially domming the whole class. We can either edit the computer's response - just cut out the line where she mentions using BDSM - or we can regenerate until we get something that goes with the mood we want to set, and the pacing we want.
I like this result better. It's more subtle, but still hints at what's to come. You might also get a response where the Coach starts dominating you right away. Or just edit the LLM's response to make that happen.
Sometimes it sucks
For the most part I like to take a hands-off approach and accept whatever twists the game throws at me, but let's remember what an LLM is. This isn't a game. We've told it that it is, so it tries to give us the next likely text in this context. The characters don't think - they aren't even distinct subroutines or game objects. They're just elements in the text that the model is trying to continue riffing on based on what we tell it.
Sometimes the pace is too fast or slow, or it adds in unnecessary commentary like "Oh! This sounds like an exciting story! Let's see what happens next!" or other goofy shit. Sometimes the responses ramble on and fucking on and it presumes actions or dialog I don't want the player to take... so I need to reign it in. I edit the responses or regenerate them until I get one I like.
Anyway, I have no idea what I'm doing, so feel free to suggest better models or prompt templates. Cydonia-22B is just about the limit of what my GPU can handle at a speed that's bearable.
I have no idea if this is interesting to anyone else but me - but if it is, download LM Studio (or a different framework) and try it yourself. Tell me how it went. Or let me know if you want to see other scenarios I've played with, like:
* Locked down during a sexy pandemic
* Psi Op with a protective squad of rugged yet sexy space marines
* Sexy House Party (of the damned)
* Accidentally assigned to a sexy all-female colony ship oh noooooo
* Your own. Personal. Holodeck. (also it is sexy)
* The Sexy Doctor Mindfuck has invented the bimbofication gun!
Okay some of those are just me riffing right now, but they'd probably work just fine.
