MarinaraSpaghetti

/

NemoMix-Unleashed-12B

like 92

What this model allowed me to do is write 20 K !!! world info (generated by koboldcpp in numbered list form (kobold ai assistant) out of a big brainstorming blob of rules i invented, changing the the world facts and rules into something completely else and changing the behavior of the LLM model only by giving it a mountain of context data. I told the kobold Ai Assistant i want to use the condensed blob as an Author's note for my chat guiding the role play and the AI assistant should formulate the rules in a way an LLM could build a world model!!! around it.

The problem is that with LLMs they are trained on either the boring real world of normal humans as a standard base or books and fiction which isn't to the liking of some ERPlers, but the average characters especially invented characters which are not in the original character cards revert to an default personality which is bland. Neighbor, Teacher, coworker appearing. The default personality is not drawn from the character card and the base info of the LLM is creating a bland npc extra character. My context based approach changed this behavior of the LLM, making additionally appearing characters more appealing for erotic role playing in particular.

For example it is a gender reversed world where females are horny like current males all the time, in our current world, ;) and men are passively low libido not interested in adult things loving romanticism and being horrified by anything too s** ual, me being the lone exception of course. The females hide their s** ual desires from the men to not agitate them, and this particular rule causes you to be able to tickle the personality of a female and show you her hidden side if you can make her trust you enough but behaving normally in society (not too horny immediately, rather trying to hide and control their horniness especially from males in this matriarchal society where alpha females go to work, are all beautiful, dont' age after around 25-30 years, and control society and males are beta males interested in gardening, soap operas and other current feminine things. Females are frustrated by beta males and me the Player is regarded as a rare and important exception because i share their female interests. For example females like video games, anime and of course s** and p***;), having big p*** collections which they hide, can't get pregnant without them willing, not having any pain when having s**, loving taboos etc...
I added everything i could think of what i would like to be different in our world in gender dynamics and human bodies and their psychological instincts and behaviors.
Essentially females behave as if they would be full of testosterone (without the male side effects of course like beard and such, i mean only the libido and interests part, more logical and s ** ual, polyamorous and bis ** ual),
and males full of estrogen (moody, emotional, romantic, naive, pure and childish, having almost no libido and interest in s** rather looking for true love and a partner to communicate with, or a strong female who goes to work and provides a safe space for her househusband and themselves believing in monogamic wholesome relationships and the physical s** part not playing a much important place in their psyche.)

I open then in Silly Tavern a character card, copy paste the first message of the card in a textfile, delete it, and paste the complete 20k rule set into the chat window. (Author's note was not big enough and i read somewhere that it is every now and then inserted as a reminder and that would be contraproductive with 20k context, System Prompt did not have enough free space but this trick with deleting the first message and replacing the first message with the 20k block works wonderfully. I pasted the first message of the character card after the 20k blob and started my role play.

And what happened was that the LLM Model changed it's complete behaviour, writing, scene orchestration and all that only because of context. LLMs by their nature are contextbased so having 131k context enabled this adaption and who knows what will happen if even more context up to 1 million like gemini or such is completely open source and we have the hardware to run it.

With this method you can get help BY the AI (for generating the world info list) FOR the AI.

I am now at 112k of 131k max context and everything still works. (^_^) 64gb ddr5-6000 ram, amd 7800x3d, rtx 4090, 24gbvram. It takes now about 450-550 second to generate a ~500 token reply but that is no problem for me because i play kingdom rush tower defense game and between levels look what was generated, or just hit the swipe button to have at least 3 diferent swipes and read them. Then i often put the multiple swipes in different text file, pick out the parts i like and use that as the final answer manually editing here and there.

Essentially what this experiment showed me is that by adding a big blob of high quality context world rules and examples of what you want to experience in your roleplay, even amateurs not knowing anything about finetuning models can steer the LLM in the desired direction and responses.

Sometimes logic errors occur of course but i suppose they would be less if we used a 70k or the 405k model of LLama3.1.
But then the context window and quantization would be so detrimental that you could not achieve what you can with this 12b Nemo model and high context.
If the 12b model here generates wrong text (confused about names, body parts, places and such) in my experience it is about only 1-2 out of 5 alternative swipes which give you bad output (logic errors or other errors you could describe as being confused.)

Essentially this model is a "State of the Art" roleplaying model and MarinaraSpaghetti can REALLY BE PROUD OF WHAT HE HAS DONE!!!
Thank you for your contribution!!!
(^_^)