Love/Hate
I love and hate how good this model is. my pc cannot handle a larger model than 12b, im using the imat q6 version of mag mell and it feels identical to the q8 version of it which is nice, less space, more speed and all.
I have been stuck with this model for a year now, its good. like, very good. i have heard of such models of similar size that people boast 'Oh its better than mag mell!" no its not. ive tried, good lord ive tried to find something better, tweaking settings, adjusting stuff for days on end, either im stupid or i just keep getting pulled back to this one. it follows my formatting instructions rather well, its a "good" model.
That being said, its not a "great" model, but its the best 12b model i can find, its creative, it doesnt get confused with formatting (i might have to tweak its output on the first couple messages but usually i dont have to)
its smart, it can surprise me from time to time with stuff.
My complaints about it is its a little too happy. there's a balance between unrealistic pessimism/anger and unrealistic happy-go-lucky and this one is jusstt a tad too happy. that being said.. hand it an asshole character and it will be an asshole, but its quick to change personalities if lets say you scare the shit out of the character, come on, people aren't so easy to change. it needs to follow through a tad.
Ive tried dans personality engine, impish bloodmoon, godslayer, nemomix unleashed, and a number of other models, im sorry, to me, mag mell is just the best out of all of the 12b models we have today. and i HAVE tried to find others that are better or as consistent at least, i usually get stuck with a scavenger hunt to find the perfect settings or templates and whatnot.
use temp 1.25, top p 1, typical sampling 1, min p 0.2, top p 1, dry base 1.75, dry allowed length 2, dry multiplier 0.85, dry is optional, top k is optional, rep penalty range/penalty is optional, anything else i believe is optional, mostly just stick with that temp, top p and min p and youll be alright. i run this at 10k context length. Oh, set include names to on.
i do have to prompt it to tone it down since it likes lengthy replies, i prefer 3 to 4 paragraphs, it pushes to 4 to 6 paragraphs but some people like lengthy replies, i use asterisks for third person action quotation marks for "dialogue" and backticks for first person perspective thought
A neat thing to note, i have instruction to use thought in the system prompt, zero examples, just this "When expressing {{char}}'s thoughts in first person, use backticks like so" and that is all, with just that, it uses them even if there are no example dialogue. this model acts like a model much larger than it. adheres beautifully to such things.
Tested with a very large character card that has changing stats, while it didnt do "math" per se like from 100 to 95 (-5), it did shift the stats in a way that makes sense, it skips the instruction to add the plus or minus and just goes straight for guessing 95. it does (like most models aside from 70b plus) like to hallucinate new stats, thats a bubble that has yet to pop for the last several years. so stat changing is yet to be perfected, this model is just better than others.
This model uses ChatML, nothing needs tweaked for that.
I can give this model an 8 out of 10 considering its size. I want to see better, I feel disappointed that models haven't really improved much over the last few years. if i had the equipment and resources id be out there making my own models, im picky. and being said picky person, Mag-Mell is my top choice for what my system can handle.
@inflatabot, good job. this IS a good model, i joke about hating it only because it truly is the best model i can find and i get frustrated when i cannot find anything new and amazing, im an explorer, i want to see improvement, better and different things, you should be proud of your work here, but its not YOUR fault that you've set the bar for others to try to match. i hope to see great things from you and others in the future.
Okay im off my soap box now, this message rant has been a long time coming.
Very well stated! I have to agree that it is now going on 2years old and by far the best model I've ever seen or found. Eventually once I get vision grafted in I don't think I'll be using any other model.
This made me really happy to read, I'm glad it's stood the test of time with you.