Nyx-RP-9B-Instruct-2608-v1: Final release version of a roleplay finetune for Qwen 3.5 9B.
"Engineered to punch above its weight."
Quick Overview:
- Way better roleplay than the base model
- Better pacing and punctuation
- Better creative writing
- More consistent interactions
- Specialized private dataset with carefully selected data was used
Quant info:
- BF16: Overkill
- Q8_0: Highest quality, mostly overkill
- Q6_K: Near lossless, recommended
- Q5_K_M: Very high quality, recommended.
- Q4_K_M: High quality, fast, recommended
- Q3_K_M: Lower quality, fastest, recommended.
Quants are all located in the repository.
I am currently also working on gemma 3 12b it finetune, but this one did surprisingly well.
Note:
I have also now released my similar gemma 3 12b it finetune, in comparison to Nyx its more creative, but Nyx is incredibly smart and follows instructions flawlessly, issue with qwen models writing dryly is less noticeable here after my observations, I'll leave a link to it here just in case: https://huggingface.co/Indexnusrefather/Erebus-RP-12B-Instruct-2608-v1
UPD: Benchmark attached, currently leading 9B and under ERP category on caliperbench.com:
I will continue gathering more high quality RP logs to use for the v2, but I'm already amazed.
v2 needs trial:
I updated my methodology and released a couple of different test versions which will later become Nyx-RP-9B-Instruct-2608-v2, I made them in an org not to dump everything into my main profile, they are named Interferons(from alpha to gamma for now) and you can find them here: https://huggingface.co/Wonderlab-Testing-Grounds
Feedback for v2s is very important, I recommend trying out the gemma version as its the most promising one to me this far, but others might be good too. First results were as expected, testing versions are testing for a reason, most of them sucked, the v1 still leading in the ERP category, now I know what approach to take though. For now I will leave this for later, I'm currently more focused on my other plans and projects.