Best RP model of all time?
After more than a week of testing, Loyal Macaroni Maid has become my all-time favorite RP model, ahead of those I considered the best, including the 20Bs. And I've even tried the latest ones. None came close.
It's one of the few who follows instructions and context well. Its style is a pleasure to read, not like some overly heavy or even repetitive models. And it's also very creative.
I can't believe it's only 7B...
And 8K context, baby!
Do you think it would be possible to correct its flaws without losing the magic of the original? (The flaws being mainly, for me: slightly and periodically (re)censored, and its tendency to speak and act for the user, no matter the instructions I give him, even 'positive' prompting.)
And perhaps extend its context even further? 16K, maybe?
I gotta +1 on this comment. Your models (even the last ones) are really impressive Sanji but the flaws are too strong. I mean, the impersonation of {{user}} is so strongly inherent of the model that you can't prompt against it. Even if I take a profund care of my character card to avoid any mention of {{user}}, it still happens too much. Don't get me wrong, I know its only 7B, and impersonation occurs with other models too, I'm aware of that.
From my experience, this is a flaw coming from the Noromaid iterations. So, in a nutshell, Creative, Stick to character, but impersonation bias so strong it keeps me away from using them as daily drivers.
Anyway, thank you for your work :).
Thanks for the feedback!
I'll try to construct an impersonation test and add it to my test suite. An extra unalignment pass might also be helpful for some instances of the minor censorship left in the model.
I have to also praise your work. This is the best model I've seen for RP, even only being 7B.
I've also used it in sillytavern to generate prompt for stable diffusion images. It generates very decent prompts for stable diffusion given you change your image generation template prompts. The model being this small also means it's possible to run everything locally (i5-12400, 32 GB RAM and RTX 4070).
I haven't had any issues with the LLM generating text for user unless I'm using instruct mode. The model seems to perform a lot better on chat mode and I'm using the "Roleplay" context template from SillyTavern.
I can agree. I personally test models and this one scored 27 when other models max at 13 points. I would love to see a merge or bigger variant.
Thanks for all of the kind words!
I can agree. I personally test models and this one scored 27 when other models max at 13 points. I would love to see a merge or bigger variant.
I saw similar results in my tests - not as strongly as 27 vs 13 but my 7B were beating the field :)
@SanjiWatsuki
Leaving my comment and two cents. I have to say for 7B it's such a great and fantastic model! I am so impressed how the result are so damn good in comparision, given that I normally use models inbetween 13B and 30B.
I would like to know, if it's normal that this and your other models are strongly adviced on consent? I tried so many options to push the behavior into a more offensive stance for the model, but it's very defensive on many topics and says "We should do it carefully" or "respectful environment rooted in trust", especially on topics like stealing something from a house, jumping from a cliff or trying to get the new king by stabstab. It's great for roleplay in almost every time, but sometimes it feels censored.
@SanjiWatsuki Your model interested me, I have never seen a model that could instantly respond to the simplest requests; even free online models cannot do it at such a speed. The worst thing is that such models can get lost in search of “that very dream model.” , and especially new users who are on the site for the first time see models they don’t need and this search takes a lot of time