Feed bag and *hicc* samperlers (i grant thee... 5 stars!)
Okay, my initial impression was "Eh, meh", some responses felt somewhat cliche. But the more i played around with it, the more it grew on me. I put it away for some time and tried out UnslopNemo v4.1 in the meantime. But it just couldn't click with me, bland prose, troubles with reasoning... Then i loaded this model back. Boom. It was like night and day. The responses felt great right off the bat. In a situation where something absolutely outrageous happens, it actually reacts realistically but vividly, adhering to the character card, and the prose is pretty good. I kept rolling with it, and it kept giving. Some responses even actually made me chuckle. But then it kept surprising me even more, and it slowly dawned upon me: Holy shit, this model COOKS!!!
For the past month, i've been a SLAVE to Lyra-Gutenberg. I tried out other models, but nothing could compare. But i wasn't happy, it was just the least sucky option... Okay, not true, it was really good, but i needed MORE. I've tried so many 12B in the past two months, it's not even funny: Chronos Gold, RPMax, Gutenbergs, StarDust, Rocinante, Celeste, Dark Planet Titan, Star Cannon, Nemomix Unleashed - that's just the ones i can remember. There were 8B models too, but i don't even want to talk about them!!!... Anyway, most of them ended up in the recycle bin on the next day, if not within one hour.
And here i find THIS. Something that steadily ticks every box for me... I'm feeling like Season 2 of Mistral Saga might be ending after all... The Gutenberg Slave arc is coming to a close... The farm has been illuminated with Violet Twilight! (mighty river, release my soul, i need to rp...)
Before writing this i was like "Hold on, wasn't this supposed to be just another one of "those" merges? Why is it so vivid and fresh all of a sudden?" And then it dawned upon me "Oh, silly me, that was StarCannon!"
So, i see it's a merge of the models you finetuned yourself, and suddenly it all falls into place. Which now makes me want to check them out as well! You did a really great job!
P.S. Actually, i might have figured out why my initial impression was mixed. I used the samplers suggested on your model card. Two of them had reasonable temp - those are the ones i tried, i didn't even touch the ones that had temp of 3 and 5 (o_O). But then i switched to my default samplers that i use for all 12B models - which are STUPID SIMPLE, but they also just work. But i think it's one specific aspect that made the difference: you see, those samplers from the model card, they use XTC. But i don't! I actually find that it tends to do more harm than good!
My STUPID SIMPLE it just works samplers look like this: temp 0.7 (the OPTIMAL for ALL 12B), minP 0.02, and DRY with default parameters (0.8 mult, 1.75 base, allowed length 2, range pen 0). That's it!! It just works!! It's so stupid simple, there's no need to even upload a json, just neutralize all samplers and dial these 3 in. Dial and forget!
With some models i also try pen rep at like 1.03-1.08, but yours doesn't need it! Nay milord, need it does not, it needs it not, not, not, doesn't need it!
BTW, i'm using Q5_K_M quant.
P.P.S Clasps hands together in a prayer "Dear LLM Santa! I dunno who Epiculous is, but they have COOKED!! Oh, let it be so they continue to cook!... Mayhaps a finetune of Mistral Small next!... Oh, so let it be, oooh so let it beee~~~"
Cheers!
I agree that this model is utterly insane, I had tried Azure Dawn and was already blown away, but this is really on another level. The character I'm making (a girl who doesn't want to talk to you) just came alive.
It seems to have a habit of resurfacing info that's just about to leave context right before it does in a coherent manner. It has excellent temporal coherency, excellent ability to switch between narration in first and third person forms. It doesn't really repeat itself, like you said, rep pen not really needed, though it may be slightly more interesting at 1.05. and just using DRY with base settings is fine. Azure Dawn handled temperatures of 2.5 fantastically, though this model is pretty much broken at that. I like it with dynamic temp 0.1-1.5.
It's very good at advancing the roleplay on its own, if a character has any incentives, it will pursue those primarily which is great. It's not easy (as the roleplayer) to make the character behave in an uncharacteristic manner. It's also crazy good at understanding implications. Normally you'd often have to make implications obvious, but not here. Like I had a whole bit with my character where she insistently reminded me how we were just friends despite her inner monologue constantly thinking about more-than-friends things and us already doing those things, so I kept increasing intimacy, saying things like "normal friends do this.". Normally, I'd expect a language model to break here and to misinterpret my character and make her shut down, but she played the undertones with me. Yes, friends kiss. Nothing abnormal about that at all. This model does that and isn't limited to one level deep.
Also using Q5_K_M, ~30K ctx.
Looking forward to Epiculous' next model(s) and wasting more of my life in the meantime.
Alexa, play despaticko.