"Until I mute the microphone (in game) My mic is physically muted, Every two seconds "YOU" is written in the chat box."

I had not thought about that, I have no idea how to recognize that the microphone is manually muted (hardware), the deactivation is by software, I think that "YOU" is noise even if it is manually muted.

"At the first launch, I shut down the progam manualy in window task. It don't want to close itself. Only happend once. Now it close as it should."

I suppose it was when you start the game, it is the moment where the model (LLAMA 2) loads and takes up a lot of RAM.

"After arround 30 min, I can't write anything in the chat box. I wait a bit for a reply from Marian, but nothing. She don't talk and I can't write neither."

There is a point where the context is unloaded to have context space again, there it takes a little longer than usual, but it shouldn't be too much... the other is that the context became saturated but it would be very strange... in tests, if I put lots of text then LLAMA breaks... but there is a condition that limits the amount of text, anyway I will check it.

Thank you very much for the feedback, the game is still in development! Greetings

Originally posted by JORGE:

... I think that "YOU" is noise even if it is manually muted.

From where? My mic can't or shouldn't produce any electric impulse. With discord, it work well. (i bought this mic over a year ago and didn't use often - Razer Kraken V3)
A bug between my software and your game perhaps?
I will try more to be more precise.

I suppose it was when you start the game, it is the moment where the model (LLAMA 2) loads and takes up a lot of RAM.

Arround 20 and 30% of my CPU and a little more than 1,2Go of ram
if i remember correctly.

There is a point where the context is unloaded to have context space again, there it takes a little longer than usual, but it shouldn't be too much... the other is that the context became saturated but it would be very strange... in tests, if I put lots of text then LLAMA breaks... but there is a condition that limits the amount of text, anyway I will check it.

I will try also again if that happend every time.

Originally posted by Lunjei:

A bug between my software and your game perhaps?

I did tests with a very old microphone that I have, it does not have a switch to put it on mute, what I did was disconnect it and see how the YOU comes out. I don't know if I should call it a bug, it's a way in which the whisper++ library that I'm using behaves hehehe.

Originally posted by Lunjei:

I will try also again if that happend every time.

Thank you very much, I really appreciate your points. Greetings!

So i have monitoring each session i did with Marian.
When i input a question or a response my CUP hit 92%, sometimes less, depend of what i'm saying and Marian respond. If i said nothing it is as i said arround 12% of cpu.
Ram is usualy stable arrond 3.6Go. (not erery time) But don't change much.

LLM seem to crash itself a bit after Marian an i are in the holographic sphere's room.
I said seems, because i try to spend more time in the lobby near Marian. even longer after she unlock the door, to see if that crash can happend there.

Do she have a scenario of something like that. In the last corridor, we can't go further. Does that trigger a crash, beacause she don't have a leading plot anymore? Her leading plot it is something like that i beleive?
1_In lobby meet with player, talking about him or her.
2_Cortana open the door
3_find a way to exit the building.

Until now each crash have happend after we arrive in the last room. (longer or sooner)

I have tried for 20 minutes for Marian to curse me... I try every trick i could imagine to make her do that.
She don't have an ego or introspection skill so of course, bashing my head on the the wall have the same result. lol
But it is fun to try to break her rule.
"... i will not try to curse, because i 'm respecful of myself... and we should focus on how to get out this building..."
How many times she said something like that... i don't remember.... :D :D :D

Thank you very much Lunjei, if you agree I will add you in the credits as: game tester.

When you refer to the crash, are you referring to what you said before Marian stopped talking?

Or is it another problem related to the video game closing due to a fatal error?

In the case of GPT NEO, it was trained with redditt and twitter. So, context has little influence on the bot's personality. So, It's easy to make the bot go dirty :p

In the case of LLAMA 2, (and LLAMA 3.1, I already managed to run it, it's pretty good...) the context does influence the personality of the bot. My context has phrases that skew his personality towards being more respectful. For example, this is a small part of the context:

"Marian has the INFJ personality. Marian is self-aware and very empathetic. Marian is sensitive to the needs of others. Marian is emotional and logical when making decisions. Marian understand the world and hope to be able to make it better. Marian is looking for deep relationships. Marian has a strong sense of intuition and emotional understanding. The most important thing in life for her is love."

Originally posted by JORGE:

Thank you very much Lunjei, if you agree I will add you in the credits as: game tester.

Not crucial, but i appreciate that gesture. So yes, thank you.

When you refer to the crash, are you referring to what you said before Marian stopped talking?
Or is it another problem related to the video game closing due to a fatal error?

Yes it was what we spoke about. Sorry didn't exactly pointed out what it was.
The game itself seem working fine.

In the case of GPT NEO, it was trained with redditt and twitter. So, context has little influence on the bot's personality. So, It's easy to make the bot go dirty :p

Indeed yes, she seem shy and reserved, but start to warm up her... Not as much i could like but it was hot. ^^

In the case of LLAMA 2, (and LLAMA 3.1, I already managed to run it, it's pretty good...) the context does influence the personality of the bot. My context has phrases that skew his personality towards being more respectful. For example, this is a small part of the context:
"Marian has the INFJ personality. Marian is....

Thank you, i understand better, that's good news.
So each episode will have the same personnality's context? Or can you change something, depend on what objectives we have to do?
Does she will remember something each new episode about what happend previously?
Do LLAMA 3.1 can keep more longer some "token" i have said?

I didn't mention about music before, i like them! About that song "More than words" someone did that for you? Suno AI?

Ok, great!

All songs were made by Suno AI, generally the words in the prompt I used were: synthwave, ethereal, energetic, female voice, something like that and of course, trying and testing and only selecting the ones I liked. The only lyrics I wrote were the song: More Than Words (https://www.youtube.com/watch?v=GXxj05KbSu8).

Regarding the problem, yes, most likely LLAMA 2 is crashing, in the new update I will add a log. I believe that with the larger context window of LLAMA 3.1 it is less likely that the error will occur.

So each episode will have the same personality's context? Or can you change something, depend on what objectives we have to do?

Yes, it is possible to introduce another context with new objectives.

Does she will remember something each new episode about what happened previously?

It is also possible.

Do LLAMA 3.1 can keep more longer some "token" I have said?

Yes, LLama 3.1 has a much larger window.

By the way, I already have LLAMA 3.1 integrated and it is far superior to LLAMA 2, it talks a lot and I don't like that so much, but it is quite good at responding.

That's all sound promessing !

... it talks a lot ....

lol, well, we will see. ^^

Good thing with a log file.