Feedback
So this model ive been testing for the past week or so since i found it trending.
Refusals: encountered only one refusal since i began testing it, so it exists but its rare
LONGG responses unless i stick this prompt somewhere in a high priority area. "Less is more, you are not writing an essay so replies should be no more than 1 paragraph long. Never assume that paragraphs should ever be longer than 1 paragraph and line breaks are strictly forbidden."
Long responses seem to be a thing with a lot of models these days but that is alright.
This model seems to like to occasionally go off on random yet small tangents as if the temperature is set a little high, its kinda rare so a re roll fixes it. i have temp set to 1.2 in all of my tests and i use ChatML for this model of course.
This model has good creativity, it can keep a roleplay flowing nicely (except when it spouts off huge responses) and is my current favorite 12b model.
Spatial awareness is decent, while not perfect its good enough. id give spatial awareness a score of 7 out of 10
intelligence. its like.. mathematics kinda, score keeping. its decent though again not perfect. ill go with 6 out of 10. but im being very unfair to it since i create a character card with complex scoring systems and precise instructions for the model to follow, hitting it with various scenarios that would adjust the score depending on what is happening. this is very hard for any model to do and so far, only 70b models can handle this type of score keeping accurately.
Moral compass. so this is what creates refusals imo. while the model is uncensored, it does have a preference to not do dangerous things and it literally tells me off on occasion. had this issue more frequently with my last favorite model (archaeo), so if i had to guess, adding a little darkness to the data might help? the model avoids natural death or dangerous scenarios, id love to see an amoral magnum model, like an "It is what it is" as in, whatever its prompted to do. ive attempted to add model instructions to make things dark but it resists pretty hard.
suggestions:
Include dark stuff to balance out its training data to make it more amoral
somehow adjust training data for less lengthy responses? it really loves to chatter.
Testing using backyardai, using quotation marks for dialogue and nothing for action and stuff, ive found using asterisks or other action markers screw with the quality a little.
Overall, this is my current go to model. im excited to see improvements
Thank you so much for your feedback! Appreciate it alot.
Amorality
Think this is possible with some other datasets that I didn't really include because of not being "Magnum"like. Which means just being generated by Claude. I've also been thinking about a plan to make models prefer darker outputs rather then Positivity in a blog I'll be writing shortly. But that does mean the overall tone might be less geared towards Claude-like prose.
Chatter
So that is a staple of Claude haha. Overly long verbose responses. Which is a reason I've been working on my own Roleplay datasets, the first model to be trained with it being my New Gemma3 12B models. They often prefer shorter responses. I think either merging with a shorter response model would be cool, But so far I haven't found another model to merge with that would be a good stand-in to make an Archaeo-V2
Intelligence
Yep, Fix for this coming soon with an RL trained version. Which should hopefully boost coherency by a mile. Just whenever I can nag some GPUs and finish cleaning data.
If you wanna push it, You can donate on my ko-fi
https://ko-fi.com/deltavector
Again thank you so much for the feedback, Its invaluable! <3