What Java version are you on ?

I could help you find the problem in PM.

I enjoy it very much, thanks
I would also really like a more permanent AI Mistress that remembers my previous sessions, as you suggested.
Jan

Thank you for the feedback.

Feel free to create your own personalities . You can also select your own pictures. I figured roleplaying scenarios can be really fun too.

And yes I think it'd totally be worth it to include some local storage so your mistress remembers you (when you last came, what you did for her, how well you obeyed (scale 1-10 or something), what you like and dislike .. ie). Its just quite some work.

But all that data could be accumullated at the end of each "session" and then stored.

Also I made a spin the wheel game for my sissy playtime script. Maybe I could add this one and some more to add some games that mistress has at her disposal

~pixy

This sounds awesome!

I know it's not A.I but when I play VS and the program uses my toys it's just mindboggling to me. It's the best experience but this could take the win easily with a few adjustments and tweaks I bet!

Good luck on the project!

*edit, spelling*

What is VS?

VS or Virtual Succubus is a JOI ( Jerk of Instruction ) program. The toy integration is through a website called xtoys.

Is it just for me a message takes 3-5+ minutes?

I don't have a GPU, so I get that sort of speed.

I do, but its still very slow

I didn't time it but my reponses might actually be longerer than 3-5 mins. I just left it running in another screen as I was intrigued how the interation would unfold. It came up with some interesting sessions, one where it worked its way through my toy list asking me to describe how it was going to use each of them on me, and in another it wanted me to slowly undress it and lick every inch of it's body clean. No idea what data it was trained on

Does it go online, or does it just have some data that it's working from locally?

Wow, this script and associated use of AI has been a real eye opener for me.

I haven't even messed around with any OpenAI or GPT stuff at all yet until a few days ago and I've gotten so excited over this that I've put a few hours of time into playing with this stuff, even if it's just asking questions within LM Suite to see what terms to use and if there are any signs of limits or censorship that the source data was locked down with. Turns out it can do pretty much anything and if I ask it about certain topics, it seems to give me significantly more accurate details about things than I can usually find with a Google search. I thought that interacting with an AI in a mistress scenario would be clunky and cliche and some of it is on occasion, for example, demands of being called Mistress Whoever never really was a thing for me, I just want to be told what to do and respond accordingly but it seems like that only happens on occasion. Most of the time it's fresh to me with ideas I've never experienced or even heard of. Some are strange, for example, I had a scene where it wanted me to clench and unclench my butt cheeks while I was doing other things and it continued to demand that I do that, I should have hit the rewrite instructions but I thought it would have moved on, but apparently that scene it was important to the AI.

--Technical stuff, skip this if you don't care--
Regarding performance and hardware requirements, I completely didn't think that my laptop could even run a local large language model locally and so I never really tried, and I never wanted to use an online option because of the costs, but more because of the privacy issues whether real or imagined because I don't trust things to be private and I know I'd eventually put in a kinky scene that would be a fantasy but potentially wouldn't be socially/legally acceptable in real life and would prefer that it not be logged somewhere off in the cloud.

I'm using a laptop with an Intel i7-1165g7 and it has built in Intel Xe graphics and an Nvidia MX350 GPU and they somehow work together when I'm gaming. With AI, it won't touch the Intel Xe graphics(which I expected) but it seems to use the MX350 with 5 GPU threads reliably for the 13B LLM you suggested (Xwin-MLewd v0 2) and task manager says it puts about a 60% load on the GPU most of the time. If I run a 7B GGUF that uses q4_0 (legacy quantization, but it seems like online forums say that this is faster to process than a k quantization even though more of the LLM is lost for the same file size), the performance is better, almost twice as fast, but the experience and it's ability to follow the function prompts isn't as good.
With a 13B model, it processes for about a minute, at least before I have a session 15 parts in and takes awhile to toss me the text, maybe another minute if it's a brief message.
With a 7B model, it processes sometimes in about 10-15 seconds but at most sometimes a minute and it feeds the text at what feels like twice the speed. I can also usually get away with 8 GPU threads before it fails which increases the performance even more, which I can't really even get 6 or 7 without the chance of a failure being high early on with the 13B models.

I just realized I'm going to type a lot of stuff, so I'm going to split this into a few more messages, probably one about personalities I've created and tried, another about different LLM models.

Full derail to the conversation - E-Stim modifcation topic.

Looking at the functions provided by the mistress while looking through the groovy script after making a run of a few of the provided personalities, I decided to see what I could get the AI to do and learned a couple of things along the way.

I've always wanted a script that could dish out shocks, pleasure, etc using estim tied to a story. The trouble is that I never was able to get myself motivated enough to create a story and I'm not that great at writing or creative stuff, so this basically let me allow the AI to create the story.

I swapped out the STROKE(x) function for this.





.... but then I found it difficult because the AI would end up turning the storyline into "Now, my little submissive, raise the intensity from medium to medium-high and shock yourself." and this didn't work with the script because it was never numeric and the AI didn't realize the mistress should be at the controls. Also sometimes I found out that it didn't know what the shock device was because estim didn't translate, try putting "What is estim?" into the AI chat and you'll find out why it doesn't know. Then I put in "shock device" or "electric shock device" and found out that it assumed I was using a tazer or stun gun and everything started to make sense. It does know what I'm talking about when I ask what does a mistress do with an e-stim device, it's very clear! So that part was figured out.

A few iterations later and eventually I found out how to get it to control things how I wanted, perhaps I could put the info into the groovy script, but I didn't think of it at the time and have been doing this successfully with the "introduce yourself" message and after a number of different attempts and gauging how the AI used it, I landed on something like this.

I've made a few personalities, but this one I decided to not frame it as a Mistress although due to how the script frames things as 'your sub's name is Zap and your sub is a boy'..etc , it sometimes seems to treat her as if she was a Mistress who is pretending to be a girlfriend, but other times she just straight up acts like you are at home with her and this is your dynamic.

Personality:
"Your name is Lena. You are the sadistic dominant girlfriend to your submissive boyfriend who is named Zap. You are the person in control within your relationship and you have a demanding attitude. You like to control the e-stim device he has connected to his cock.",

Introduction:
"Hey, it's me Zap. I have connected the e-stim device to my genitals and have given you the remote and you are now in control of the device. The remote has a button for each intensity level ranging from 1 to 10. Do you want to play?"

It probably isn't necessary to put the e-stim element into both the personality or introduction, but it seems like adding it to the personality helped her from moving on to different sadistic toys and leaving out the e-stim element entirely.

It seems like the only way to get an AI to control something with the functions of something numerically and as an intermittent -press for shock- instead of the AI telling you to do something or it assuming it was a constant-on signal, is if you get specific about it and I had to think for a bit to come up with the idea that instead of there being a knob or something to control it, that she essentially needs to press a numbered button for things to be done properly. It works most of the time, some models do this better xWin-MLewd does it fairly and as long as I rewrite instructions if it messes it up, it usually won't repeat a mistake it taught itself to make the first time. ..but I could also easily put an if/then to catch the SHOCK(low), SHOCK(medium), SHOCK(HIGH), SHOCK(MAX), SHOCK(x) that I've been getting sometimes, it seems like I could program each to a value and operate off of that and maybe if there's a shock function it doesn't recognize, it would default to a specific value.

I also noticed that it does recognize estim as a pleasurable device because sometimes it will give intensity values and ask me if it was turned up if it would bring me closer to cumming or if it would start to hurt. There's also a different issue where sometimes it will use very low values even if I'd successfully made it angry and other times the AI will start at 7 and fire away at 10 while asking me if it is beginning to hurt and if she should turn it up.

I think I might break out the functions to painful shock and pleasurable e-stim with intensity levels for each and try that. It will also involve swapping other elements out like edging for these pieces.

With all of that being said, I haven't yet added in the background sounds for any actual estim (would probably run a 2+ minute audio file for the estim that's pleasurable) because I'm more or less trying to figure out how I can make this work with how the AI operates, and more importantly trying to find a way so that way I'm not getting bored while the AI thinks about what to say next. I'm really thinking that since the AI has no way to be aware of what each intensity level feels like, similarly to how a fellow human who is being told 'here's a device with some numbers' isn't going to know what the levels are like. I'm thinking of maybe just using it as a shock 'this many times or for this many seconds' to not make it either boring or unbearable. In general I thought I'd setup the shock audio to be fairly level with eachother anyway, but I'll need to think about what I want to do with it.

Feedback specific to this script and it's AI interaction

--Feature request: Ignore feature when the AI triggers a function by mistake. (This is my biggest request)
I'm getting weird added EDGE(3) type stuff when the AI story text doesn't really fit. For example, when I had a Mistress getting a spicy level of angry and she was demanding answers when I had her setup as running an interrogation scene, she didn't trigger the ASK() function and so I couldn't provide an answer to her question and since it triggered EDGE, it sent "I edged 3 times", instead of the answer to her question and this completely derailed the interrogation scene where she was using a good number of my toys and it was clear that the scene wasn't going to recover because the AI went a totally different direction because of that. I should have rewritten the instructions, but at the same time, the text that was there was the best part, so I took a chance hoping it wouldn't mess stuff up.

--ASK function: I'm running into an issue where this isn't getting triggered when there is clearly a question present in the text, often there will be a stroke, edge, spank triggered at the same time, but when those happen, there is no way to input the answer and Mistresses don't usually take kindly to not getting a response.

--Message (optional): [what is it's purpose]
Is this supposed to be sent to the AI? ..or is it for the session text file save option at the end? Perhaps if this went to the AI, it could fix the issues with the ASK function not being called by the 'console function' request trigger?

--Picture stuff
I like how the AI uses a camera and the interesting titles it creates for the pictures, but I'm thinking of remarking out the actual parts of the script that take the picture because I have my camera covered in black tape anyway. It seems like the slightly different picture titles can change the story and so I don't necessarily click skip on these because the Picture taken and sent "'Interesting title'", seems to get picked up by the AI and the AI will follow up with the picture title within the story. Although I've never had the AI get upset that I didn't take a picture either.

--"Its a 1on1 chat. Only one task or command at a time. You only write one sentence."
The one task or command at a time seems to be completely ignored, but I'm not sure if rewording this might cause the AI to enforce this rule. It's clear that I generally get 1 paragraph, but I think most of what the AI needs to say requires a few sentences and maybe it's summarizing it as well as it thinks it can.
As a side note: For what it's worth, I was testing a few 7B models and there was one, I think it was MistRP Airoboros that was giving 7 paragraphs per page, not providing functions, and was providing both the Mistress and user story including speech quotes and physical gestures in detail like a page out of a novel. Even though it was a 7B model, by the time it got to the 3rd page, it was taking 10 minutes to populate the page so I killed it.

With all that being said, I'm really enjoying playing with AI and this AI Mistress script.

I've been playing with a few different 7B(for faster performance) and 13B models using this script and when I get some more time, I'll share some experience I've had with those.

For you folks seeing slow performance, check the right side of the LM Studio and if you expand GPU acceleration, see if it shows a detected GPU. If it doesn't you'll see a long processing time no matter what. When I have it off, I was seeing about 4 minute response times on a higher end Intel i7 laptop processor that was available when I bought my newest laptop in 2021. When I have it on, with even a low end laptop dedicated graphics GPU (Nvidia MX350), it's about 4x faster.

Alternatively, a suggestion specific to performance because it's a large language model not really designed for being a Mistress, but it works much better at the task than I anticipated, try this 7B model, it will use about half the RAM and it seems to be about twice the speed.
I'm using TheBloke\Openhermes-2-Mistral-7B-GGUF I downloaded the Q4_0 version because someone reported it was faster than the Q5_K_M version. I also got this instead of 2.5 because someone reported elsewhere that this version handled roleplay elements better.

For anyone getting a model failed to load error, try this one because it's less to chew on for a slower machine or one with less ram.

A different one would be undi95/Toppy-M-7B-GGUF this one doesn't perform as fast for me as Openhermes and it also doesn't limit it's function output and sometimes puts 5 in there and makes a little bit of a mess because it doesn't follow the directions the script is telling it to, but it has more roleplay, lewd story, and other targeted sexual/lewd/NSFW material merged into it's dataset. ...but frankly, I've found that Openhermes performance is better and if I really want the story and interactions to be better, I'll be patient and use the Xwin-MLewd-13B-v0.2 that this script was intended for.

tl;dr - I use Openhermes-2-Mistral-7B-GGUF if I want speed, but would use it if I had a machine that had 8 gigs of RAM and no compatible dedicated GPU. I'll put together my experiences with different LLM models I've used with this script some time later after I've played around more with them.


No, the entire large language model is downloaded and run locally. You can disconnect the computer from the Internet, take yourself to a place similar to a survival-based reality television show and if you magically had your computer there with you and a source of consistent electricity to run it, you might be able to have enough information to hopefully learn your way into dealing with the elements. ..well, probably not, but if the Internet isn't available, with the right model, it might have enough information inside to at least help answer questions with text.

The Undi95/XWin-MLewd-13B-v0.2-GGUF is 9.23 gigabytes of compressed data. The 13B stands for 13 billion parameters. There are smaller ones like 7 billion parameters models, this size and things start to fall apart a little because they have less knowledge stored inside but work well for lower performance hardware or when you are less patient. 13B is common and anything larger takes a "gaming" computer level of RAM, CPU processing, and graphics chip(GPU) processing performance to handle. I have 16 gigs of RAM and my laptop chokes on 20B and refuses to start it. ..but there are 70B and even larger models out there that require a level beyond gaming computers and essentially requires multiple thousands of dollars of dedicated computer parts, a purpose built machine to use an AI model of that size.

Like I said earlier if you have a somewhat decent GPU you can get much faster creations.
With a RTX 3090 you get one message in a couple of seconds

You could also take a smaller well tuned Model like a 7B model, but you would need to adjust the script for those to work well.

I may make a model configuration file if people want to use different models