Erebus-RP-12B-Instruct-2608-v1: Complex Finetune of Gemma 3 12b it Aimed at Enhancing Roleplay and Creative Writing.
Erebus-RP-12B-Instruct-2608-v1
"Specialized dataset was used to aggressively make the model roleplay close to how bigger models do, resulting in longer messages and better track of the story."
Quick Overview:
Better roleplay than the base:
Model was trained on carefully selected number of high quality chat logs, filtered only for long term conversations with proper assistant turns.
Entire process was carefully controlled by me to ensure that model can change its writing style without overcooking, this version is a result of repeated attempts until I finally found the right setup.
Reduced refusals due to dataset containing a number of explicit logs.
Quants(this time I will release safetensors and quants in different repos for more convenience):
BF16: Overkill
Q8_0: Highest quality, still overkill
Q6_K: Extremely high quality, near lossless
Q5_K_M: Very high quality, fast, recommended.
Q4_K_M: High quality, very fast, saves a lot of space, recommended.
Q3_K_M: Lower quality, fastest.
Note:
This model turned out pretty well, Nyx is more intelligent and better and instruction following, Erebus is more creative.
I didn't select gemma 4 12b as the base because it was a hell to work with, and was in my observations way more heavily RLed than the gemma 3 12b it, so I took the older generation as the base.
Next I'll probably work on finetuning Mellum 2 12B A2.5B Instruct, also might turn my attention back to Ministral 3 2512. either 3B or 8B I don't know yet,
