Best so far
I was about to give up on qwen 27b finetunes, there was always some issue for me. This one is good from my testing. It doesnt have the repition issues or logic errors that i saw in other finetunes, though i havent tried it with long context yet. Its very creative and has good writting style. I use the Q8 with no thinking.
Parameters i found to work with no thinking
Temp: 0.7-0.9
Min P: 0.05-0.1
Top P: 0.95
DRY 0.8, 1.75, 4
At first i was having issues where it would confuse pronouns like "you, me, they, etc" saying the wrong one, but then i changed my system prompt from telling the model its roleplaying, to telling its running a world simulation as a narrator, which seems to have fixed it somehow. could have been a coincidence.
I will test it with longer context later and tell you how it does.
Edit: i want to add that i found you can increase the temp even more, depending on what character card you use. It works well at 0.9 too.
+1
I've only started playing with this one, but so far so good. Noticeably more in tune with my different chatlogs on the first few rerolls than previous fully released versions.
(also better autonomous toolcall, while I know it's not a focus of yours obviously, I appreciate it's still there).
Thanks for the feedback!
These two use a new RP reasoning dataset instead of one of my old RP datasets (Version two has one with the reasoning cleaned a bit) and Qwen seemed to take it well.
I just need to retest bluestar v2 and confirm if this is better or worse in my tests and will then likely release it.
I noticed some duplication of content during answers. Like it'd write 2 paragraphs (surprisingly well written too, it's been a while since i've seen a model adhere to my character sheets like that), and then repeat the same 2 paragraphs. Not the biggest deal and quite uncommon, but worth mentioning. I think it's mostly a sampler related thing, but I'm no fine-tuner :D
The reasoning is pretty good, it makes quite the difference. I feel like Q 3.5 is really relient on reasoning, any form of reasoning, to be present. The moment it's removed, it's suddenly a lot worse than other models in that size range.
Anyway good job, and good luck :)
So, i just tried gemma 4 31b and i think maybe your efforts may be better served on that... its really good for roleplay. The base model with no thinking has none of the issues qwen 3.5 has. Also it feels like it has that extra layer of personality that 70b models have. Just wanted to let you know 😁
g4 is definitely on the radar, my own testing of it has been very positive. Just waiting on support to improve and some dataset stuff
Yeah, been playing with it. It's still really fucking annoying to use. Llama.cpp will crash every few messages and use tons of VRAM. LMStudio is as useless as it's usually is. I got it to behave properly on the latest KoboldCpp but you need to enable swa, disable context shift and use fast forwarding in the settings to make it work well (and not use stupid amounts of VRAM). At least in text completion mode, I know there are chat/jinja issues as well.
Base model works surprisingly okay, Heretic versions are way too enthusiastic about going against whatever ethic is built into the character sheet as usual, but beyond that, it's really good out of the box and the thinking ain't massive (nor too complicated to emulate). Can't wait to see your more "willful" tunes with it.