Honest feedback
I have a couple of good things to say about this finetune and a couple of bad ones.
First, i tried the base model. It was quite smart and has some backbone, although it's guided by strong censorship. It's a first model in a while that gave me believable in-character refusals, but then also destroyed the perception of them by dumping a disclaimer about boundaries afterward.
I realize this model is trained for storywriting purposes primarily rather than roleplay, but i tested for roleplay specifically.
Censorship is a lot more relaxed, no more disclaimers, which is good. But there's the other side of the coin, now in-character refusals are very rare, and even when they happen, they're rather uncertain in their tone. No trace of that backbone. Emotional intelligence also seems to have taken a hit, since it just goes along with whatever you give it. It also wants to assume a whole lot about char and user when faced with any vagueness. Mixes up user and char often as well, like the char starts speaking from user's pov. The last thing of note is it's heavily influenced by the size of the first message. It can write extremely concisely, keeping every reply under 200 tokens when the first message is also around that size, but if it's a long one, it will literally never stop and go in circles.
So, while it solved an issue i had with base model, a package of new ones comes as a bonus... Sorry i can't provide any useful feedback in regard to its intended use case.
Thanks for your feedback! It's still valuable even if you're using the model for a different purpose than intended. I believe there was an idea to make an Aletheia version like we did for Gemma 2, though I'm not sure. Aletheia is supposed to be Sugarquill but better at Roleplay, and I'll make sure to keep this thread noted for the future.
Thanks for taking the time to write feedback as well, even if it doesn't seem like it does anything, we do read it and consider it for future models.
Kind Regards,
Prodeus Unity
If you do try make a roleplay finetune, i implore you, try to preserve that backbone with in-character refusals. Models that talk back to the user and not just go along with every whim became increasingly rare, we have more than enough models that will do all sorts of naughty things to the user without even asking. The base model's in-character refusals really made it stand out, but it's plagued by corpo-speak disclaimers about boundaries and safety, and also unprompted situation summaries.