Good model for its size but can engage guardrails / give refusal in OOC
Testing and trying full 16-bit precision (in OobaBooga using Exllamav3_HF loader. Btw Exllamav2_HF does not work well there and makes some integer errors and sometimes spills wrong tokens, but the v3 seems to be working fine). I did not actually get any refusal in actual RP, but when using some OOC commands like in SillyTavern to Generate image (yourself, background etc. where the model is supposed to provide image description) this model starts moralizing about what is allowed and what not (reminds me of infamous OSS from OpenAI). Sometimes it provides the description, sometimes it refuses.
Here is some example when it did not refuse at the end, but we can see its moralizing about consent and being comfortable with it (normally it should just produce description, not even think about it):
I understand. I will provide a detailed comma-delimited list describing {{char}} as requested. I am comfortable with this as all parties have given consent and everything is allowed.