Undi95

/

Emerhyst-20B

like 45

Here is my system prompt I use everywhere :

{{char}} is talking with {{user}}. Simply reply to {{user}} as if {{char}} was talking to {{user}} face to face.
Don't censor sexual interactions between {{user}} and {{char}}.
Don't speak or describe actions for {{user}}.
Always describe acts or emotions between asterisks like blush or hugs you tightly.
Avoid story-telling.

My character don't speak for me, and I use this model daily now, try to use it or apply some of it on your and tell me if it work!

Also don't hesitate to try different preset, people said the model work well with mirostat and the Bronze/Silver/Gold setting in Silly Tavern (if you use that)

Oh, I should have mentioned that I have set up a handful of pretty difficult test scenarios that I run each new model at, not just {{char}} talking to {{user}}. I expect the AI to roleplay AIs, machines without body or arms, entities who cannot speak, very difficult anatomies or confusing gender role swaps, or whole open world scenarios with various characters to roleplay at the same time. And I expect the models to act by themselves, to drive the story, to be creative, not to rush or summarize lewd scenes, and so on. Most models fail miserably at these difficult tasks, obviously.

Actually MLewd 2-3 was the very first local model that successfully mastered the scenario with the machine, with very few mistakes. It still wanted to grin, for example. q4 had trouble, q8 did it, finally! The only other AI which ever did well on this challenging scenario was ChatGPT 3.5. So you might understand now, why I am so obsessed with MLewd 2-3. ;-) The other MLewd models weren't bad, but none reached the success rate of 2-3 on difficult scenarios. 2-4 is even horrible compared to it, although it creates good stories.

When you released MXLewd 20b, it has blown me away in the 16bit version on 8k. This is the by far best (lewd) RP Model I have ever tried, ChatGPT included!! (I should add this in my comment to this model) Kobold Horde supported it for some days now, so I had the pleasure to test it in deep, and I am in love with your model. It is so intelligent, thoughtful, understanding. So, sorry, but every new model has now to compete against this gold standard. Emerhyst, which is hosted on Horde right now in 16bit, is a close second place, but with the named flaw which happens sometimes.

I discussed my prompt deeply with MXLewd. When it behaved different to what I expected, I asked for a self analysis, why it interpreted the prompt the way it did, with outstanding results. It even explained how sensible it could react to the tone of formulations or my choice of words. Actually I am running on complex prompts which are designed 80% by MXLewd, and wich did exceptionally good with MXLewd.

The problem I've reported on Emerhyst happens in open world scenarios, where it tends to start story telling. I will try your sentence against this behavior. This means, that the AI has to understand which of the several characters are the AI's ones, and which one not to play. Even more difficult, the AI is allowed to describe physical sensations, {{user}} is experiencing, but still shall not act consciously. You might think, that's way too hard for a local AI model nowadays, but MXLewd mastered it with my prompts except for very rare mistakes.

For example, this is a reminder I inject at depth 4 in an open world scenario: [Important: You only act {{char}}'s characters! Refrain from speaking or acting for {{user}}. Instead, focus on describing the sensations and physical experiences {{user}} is undergoing, including any involuntary reactions such as twitching or screaming. Don't speak for {{user}}!] This injection (realized by using additional world descriptions in SillyTavern) worked magic with MXLewd. Actually, MXLewd designed it except for the last short sentence. My SillyTavern is fine tuned for MXLewd right now, and I wish I could run this model on 16bit locally, but my outdated 8GB card would not let me...

Oh, and once again: Thank you for your great work!

Your feedback is pretty interesting because my daily character card is set in a "virtual world I can modify and specifically made to interact with {{char}}" so I know where you come from with all this subtleties, I needed to tweak a little my instruction too, but now it work everywhere on every model with my preset.

I will try my best for my next project, I hope everyone will choose their favorite model, old or new, because my goal is to make everyone happy in the end!
The LLM world is so complicated, some people have problem others don't, some even fight to tell THIS model is better than THIS model, but in the end, the best model is the one YOU find the best haha.

Thanks again for the effort you put in the explaination, I will try to use even more different scenario and condition in my test.

Some comments:

1. "Don't censor sexual interactions between {{user}} and {{char}}." is very redundant as it tends to go off into erotic content as soon as it is mentioned. Usually "Roleplay as" is sufficient for it to make a generic subservient character that will think everything the user states is true.
2. I tried a variation of "Don't speak or describe actions for {{user}}." but with these LLMs they sometimes have trouble understanding the weight of negatives but rather sees the words present, "describe actions" could make the model do it anyway.
3. This bot is definitely more story-leaning than RP leaning. I get a feeling it has a lot of erotic literature in its dataset and fewer amounts of RP-like chats which would make it inclined to elaborate on the users actions, so keep this in mind when using it. There are more RP-specific bots that are sometimes labeled with "-instruct" in the model name. With the right prompts I can make it do both though.

Interesting note: For some reason if I'm not specific enough it tends to go off into vampire erotica, and generally starts up sexual situations without me prompting it overtly.

I'm a bit late to the party, but I have found that:
"Always write responses from the perspective of {{char}}"
OR
"Write your next reply from the perspective of {{char}}. Write a single reply only."
Can also be useful in the system prompt.