Excellent in my (limited) use-case
Hello! I've been using TheBloke's GPTQ version of this for a (non-erotic, but with other explicit themes) cyberpunk setting, and I've tried several other and larger models, but none seem as coherent and still creatively descriptive as this one.
It needs little persuasion to make characters antagonistic and sticks to given character traits even in multi-character chats, and even extending the GPTQ-version to 8k context it seems to recall things from any position in the context better than even Psyfighter and Tiefighter (which doesn't make sense to me). It's also MUCH less prone to repetition compared to other similarly coherent models.
I'm a little confused why this model hasn't gotten more attention, but maybe it's just particularly suited for my setting? The two mistakes it seems to do semi-consistently is adopt kemonomimi traits from other characters, and write excessively long responses even when discouraged. It also uses some phrases excessively ("tension in the air", "room/air hums with X", "fingers dance in the air", "(warehouse) at the edge of town"...) but a single mention of avoiding those phrases in the context fixes it. Also rarely it gets flirty with characters that doesn't make sense, but better that than being unable to flirt at all.
Anyway, I've been checking your releases now and then to see if you've made any similar models, but it seems you've been exploring merging different base models than what you used in this one, so I'm finally commenting both to show my appreciation and in hope that you'll extend or experiment with models similar to this one. The only other model that's performing similarly well in my case is Chronomaid-Storytelling-13b, so I'm assuming chronos is part of the "winning formula". Would be really interesting to see how this model would change if its parameter size was larger and/or if it had a more varied training set.
Hello, thanks for the feedback!
Chronos is an essential element but it also use a lot of sentence we are bored with AI, sentences we see everywhere, that's why we don't use it a lot anymore in merge. It's also severely outdated.
If it's good for your usage tho, I'm happy that it can fulfill your goal!
Hello! I've been using TheBloke's GPTQ version of this for a (non-erotic, but with other explicit themes) cyberpunk setting, and I've tried several other and larger models, but none seem as coherent and still creatively descriptive as this one.
It needs little persuasion to make characters antagonistic and sticks to given character traits even in multi-character chats, and even extending the GPTQ-version to 8k context it seems to recall things from any position in the context better than even Psyfighter and Tiefighter (which doesn't make sense to me). It's also MUCH less prone to repetition compared to other similarly coherent models.
I'm a little confused why this model hasn't gotten more attention, but maybe it's just particularly suited for my setting? The two mistakes it seems to do semi-consistently is adopt kemonomimi traits from other characters, and write excessively long responses even when discouraged. It also uses some phrases excessively ("tension in the air", "room/air hums with X", "fingers dance in the air", "(warehouse) at the edge of town"...) but a single mention of avoiding those phrases in the context fixes it. Also rarely it gets flirty with characters that doesn't make sense, but better that than being unable to flirt at all.
Anyway, I've been checking your releases now and then to see if you've made any similar models, but it seems you've been exploring merging different base models than what you used in this one, so I'm finally commenting both to show my appreciation and in hope that you'll extend or experiment with models similar to this one. The only other model that's performing similarly well in my case is Chronomaid-Storytelling-13b, so I'm assuming chronos is part of the "winning formula". Would be really interesting to see how this model would change if its parameter size was larger and/or if it had a more varied training set.
Curiously this echoes my experience as well. I was running a GPTQ version of this as well, but later upgraded to a larger 8bpw-h8-exl2 after upgrading my GPU. This responds in role-play better than a lot of "modern" and larger models (I tried various 8x7bMixtral models, in 3.5-3.75bpw exl2 that can run on a 4090), with the only drawback with this being its limited context size (I also doubled it to 8K). I'll also be sure to also give the Chronomaid model you mentioned a try.
Just curious, are there any new models during the past month or so that you came across and regarded as excellent for your use case?
Yes actually, after doing some more searching and trying, I found that chargoddard/Chronorctypus-Limarobormes-13b and chargoddard/storytime-13b, which is based on the former. I use TheBloke's GPTQ-versions of both, however. They give very similar responses to Chronomaid from my experience, but sometimes I prefer the response of one over the others.
EDIT: Also Estopia/EstopianMaid, Utopia/UtopiaXL and Psyfighter are sometimes useful. The former ones for creativity/flavor and Psyfighter for following specific instructions/comprehension. Still, this model seems the best balanced of those.