Don't worry i'm already on that. :p

The game is currently configured only for Google Gemma 3 models

So my others minds gguf will not work?
I can select them in the settings but do not seem to be used. The memory file is always title with Gemma's mind. And is behaviour is not what i know.

Did the IA to walk arround the house have an effect on the way it answer?
I don't have done all test i wanted yet. (i will detail that asap)

Thanks for the uncensored link. I will try them, hope they are better than the Gemma by default.

Don't worry i'm already on that. :p

Great! Thanks!

So my others minds gguf will not work?

The GGFU models I've used are Llama models. The game can load them, but since it uses the Gemma template (chat format), the results are undefined. I've considered supporting both (LlaMa and Gemma), but Gemma is significantly superior in my tests. As a software developer, I adhere to the KISS (Keep It Simple, Stupid!) philosophy and will only allow Gemma.

Did the IA to walk arround the house have an effect on the way it answer?

No, at first it's in story mode. Once the memory file is saved and the door-opening event occurs, random movements around the house are loaded in subsequent sessions.

I will try them, hope they are better than the Gemma by default.

Thanks! Let me know, please!

So... here what i thought of this Gemma.

Uncensored version is indeed as it is and i didn't push it yet to the very last corner, but in general i can do nasty things with it.
But my problem with that model is the tunning, is not what i want. As far i can see, there is no specific tune with it.
I'm sure in Hugginface community someone will do that sooner. ^^
(i will go check it bytheway)

"superior capacities"
I don't really agree with you. His suporiority is less obvious than Gptneo against Llma.
It have indeed a slice better (or fluent) speaking (at least in french), but hit the wall yet enought with double thinking.
And do too much mistake to be tag "superior".
Concept "me" "you" is sometimes reverse. it use word or even rarely a sentence in different langage, don't happend often but it happend.
Word with gender are also difficult for him.
Funny thought, sometimes it remind me Gptneo when we started here.

I'm not sure it is useful to have the "Sampler" slide in the setting.
Even few points of difference (from the default set) give awfull result. It become indeed too repetitive or have say thing with zero sense. lol
I'm pretty sure we have spoken of that one day, but what it is the diffenre between "Mirostat-2" and "top k"

Can i make a wish?
Could we have aside of your story game, a blank mode. no visual, no story, a "pure" Llm with all vocal features you have done, only driven by our prompt.

(You have certainly see that already, the chatbox have a visual bug when we wrote a too long text.)

Thanks so much for the help and the report!

The uncensored version is indeed what it is, and I didn't push it yet to the very last corner, but in general, I can do nasty things with it.

Great!

I don't really agree with you. His superiority is less obvious than Gptneo versus Llma.

I agree.

It's indeed a better (or fluent) speaking (at least in French), [...]
And I do too much mistakenly to be labeled "superior." [...] Funny thought, sometimes it reminds me of Gptneo when we started here.

Gptneo was a crazy guy Lol

I've tried it the most in Spanish; my experience is that it's more fluid, more concrete, and concise in its statements. Gemma also manages to understand the context better. Gemma makes the mistakes you mention, but to a lesser extent than Llama.

On the lmarena.ai page (where the models compete) in the Leaderboard for the language category (chatbots), these are the current positions:

12: Gemma-3-27B-it
17: Gemma-3-12B-it
33: Gemma-3-4B-it

98: Meta-Llama-3.1-8B-Instruct

I must say that internally I'm restricting Gemma's initial batch size, which is surely affecting its performance. However, with larger batch sizes, memory usage is too high. I think I'll add this as a parameter for the next update.

I'm not sure it's useful to have the "Sampler" slider in the settings.
Even a few points of difference (from the default set) gives awful results. It becomes too repetitive, or has zero meaning. lol
[...]"Mirostat-2" and "top k"?

The model gives you a set of words that could continue the sentence, for example:

I love the {life, sea, your, idea,... etc.}

The sampler chooses which of those words to include. Top K has static parameters, and Mirostat adapts itself; these are internal algorithm parameters.

I must confess that I tested the thresholds in the slider with Llama and didn't have time to test them with Gemma. It's still pending, and thanks for the note.

Can I make a wish?

Sure

Could we have aside your story game, a blank mode. No visual, no story, a "pure" Llm with all the vocal features you have done, driven only by our prompt.

Do you mean just putting Marian's model on a white background, without her moving in space?

You've certainly seen that already; the chatbox had a visual bug when we wrote a too-long text.

No, I haven't noticed it. What's wrong?

Thank you so much, Lunrei, for your valuable comments.

About chatbox bug.
I think it happend when our 2 scrolls bar appear. vertical and horizontal. Also maybe only with horizontal.
Write a long texte and it should happend.
note : I don't play it in fullscreen, windowed and not full size. (My screen is big enought, a fullsreen mode is not needed.)
*erratum*
I change some graphic settings (windows parameters) and i'm windowed full size and it did"nt have that problem.

Do you mean just putting Marian's model on a white background, without her moving in space?

Not even Marian, no 3D model or background. Only chatbox and the vocal features with our prompt and mind model.

It's still pending, and thanks for the note.

Ha ok. thanks to remind me. ^^

I did pickup some other mind model this afternoon, including amoral-gemma3-27B-v2-Q6_K.
just to see. I don't think this one will be very different.
Ho gosh! how wrong i was to think that. My first test it F...G GOOD. I didn't laught that much with an LLM since a long time.
As i said, it is not how much dumb or smart it work, it's witch vocabulary it use who have my interest.
the uncensored version does nasty things but keep a neutral and analytic tune. Not very funny when you want salty and kinky. :p

About chatbox bug. [...] Write a long texte and it should happen.
[...] *erratum* I change some graphic settings (windows parameters) and i'm windowed full size and it didn't have that problem.

OK, let me check. Thanks.

Not even Marian, no 3D model or background. Only chatbox and the vocal features with our prompt and mind model.

I'd been thinking about a similar project. I'm taking the MTW core and removing UE5, to create a visual novel with realistic images without having to load UE5. It will be very light and similar to what you're talking about.

Link: https://www.youtube.com/watch?v=yAkYsTERSNw&ab_channel=SoulShell

Of course, I'll continue with MTW; in fact, they'll share the same software core.

As I said, it's not how much dumb or smart it works, it's the vocabulary it uses who has my interest. The uncensored version does nasty things but keeps a neutral and analytic tune. Not very funny when you want salty and kinky. :p

Yes, agree, maybe a model will come out that's really tailored to more spicy conversations.


Thanks a lot!