Harbinger V2 with less repetition would go hard
I used this model extensively in AI Dungeon and it's still my go-to for local. I've tried so many other 24B storytelling models, and I always go back to Harbinger when they start failing me.
It's fairly smart, with decent plot adherence and understanding. Good memory, respects author notes/instructions, and can write both calm and tense moments. It drives the plot forward in a logical path.
But my god is it repetitive, even with high temp or repetition penalty. During dialogue, it knows like a grand total of 7 prose lines (smirks, exhales softly, leans back, grip tightens, lifts cup, etc...). If not for that, this model would be the perfect RP model, because I am sick of characters smirking at every single thing or always exhaling before speaking. It also regularly shies away from character deaths despite saying otherwise.
I don't know if Latitude knows how amazing this model is, but a refined version of harbinger could easily be the best storytelling model if they fix the repetition.
@aer32
If you are running these local then I'd point you to using one of the backends that supports string bans - I've been maining Muse-12B-Q8 as my local core since about a week after it came out and it just keeps getting better. There are slopstrings in all of these, but the use of a phrase ban to (for example) completely block strings like (smirks, exhales softly, leans back, grip tightens, lifts cup, etc...) is like night and day.
Saw this thread while I was doing a top-of-year cleanup - retrying these non-Muse Latitude models now that KoboldCPP has both phrase ban and Adaptive-P sampling.
@jebcarter
How would you say harbinger, muse and other latitude/gryphe models are at 12B compared to 24B? I'm hoping to test these soon but my PC is bogged down with a long merge right now.
@Naphula I think Muse is still the winner, even above the 24Bs. Of the others, Hearthfire is the most interesting and stable, offering more varied writing patterns and some better brains...sorta. The two Wayfarers had interesting but not very varied writing, so they got tired out kinda quickly. I like Harbinger but I think what it does Muse does better.
The others will have flashes of good prose, but then will trip over weird logical/positional/costuming elements that Muse handles well. If you are looking for merge fodder, I am really enjoying SicariusSicariiStuff/Angelic_Eclipse_12B - its got some pushback/refusal in it, but it handles spatial and emotional logic wonderfully and is a good switchout model with Muse to goose a story for a couple turns. Prose could be a little brighter but I've been doodling some merge recipes to gently add some of that to Muse.
I have been kinda one-note on Muse for like a year plus now but I am actively trying other models and putting them through their paces but I'm also style over all and the models trained off of Nemo-Base are still the champs for that. My opinion probably differs in an environment without min-p/nsigma/phrase ban or adaptive-p available, and I have about fifty RAG'd books in my style injection heap that get inserted as cut-ups shape the output.