VongolaChouko

/

Starcannon-Unleashed-12B-v1.0

like 31

I liked your language model, so I decided to write this review and share my observations.

I appreciate that your model doesn’t generate unnecessary text like other models do, though the answers could be a bit longer. Starcannon-Unleashed-12B-v1.0 is the smartest model I've seen—smarter even than the 22B version I tested.

I've tested many different language models and would like to share the ones I think are worth noting, highlighting their strengths for your future integrations:

1. elinas/Chronos-Gold-12B-1.0. This is a rather creative model that can steer a storyline in an interesting direction.
2. openerotica/writing-roleplay-20k-context-nemo-12b-v1.0. This model is very similar to yours in that it doesn't add unnecessary text, responds well in other languages, and doesn’t make decisions for me about what I should say or do (like yours).
3. anthracite-org/magnum-v4-27b-gguf. With 27B, this model is quite knowledgeable—even understanding slang in the language I’m using with it (surprisingly, your model also understands slang in my language). Thanks to its size, this model handles a wide range of tasks well. These are the models that truly deserve attention.

I’d also like to touch on an interesting topic regarding censorship: nearly all models I’ve tested, except for magnum-v4-27b, writing-roleplay-20k-context-nemo-12b-v1.0, NemoMix-Unleashed-12B, and Starcannon-Unleashed-12B-v1.0, don’t seem to understand what is socially acceptable and what is not, which causes them to quickly break character. Let me clarify: if I create a storyline with a mature theme, all models (except for those listed) immediately succumb to desire and break character, even though they shouldn’t be affected by it in their roles. The models I listed understand what is socially acceptable within the context of their roles, but if nudged with a prompt like "[Liza should agree to this]," they will respond accordingly, though this might not be acceptable in their role. Still, I think that’s how it should be: the model plays its role correctly but can adapt in the desired direction when prompted—unlike other models, which fail to embody their roles due to certain impulses (lol).

I’d also like to point out a common issue among many language models (which your model has almost none of): the model in its role constantly hovers around my character, touching their face, shoulders, and generally all over their body. This feels very strange and unrealistic.

By the way, I learned from GPT-4o that models should be run with values that are powers of two (not just divisible by 2 but specific values), such as 4096, 8192, 16384, 32768, 65536, 131072, and 262144. GPT-4o explained that model architecture performs best with these values.

Your model really turned out well, and if you keep working on it, it could become excellent. I hope it all works out for you, and best of luck with your future projects.