ConicCat/NemoPlotter-12B

Model Details

A consumer hardware friendly hybrid reasoning finetune of Arcee's Qwen3 distill onto Mistral Nemotron 12B.

NemoPlotter mainly focuses on improving open ended reasoning / plotting for chatting, creative writing, and roleplay.

A good amount of complex instruction following and chatbot data was included as well for strong adherence to complex system prompts.

Chat Template

Qwen3. Reasoning can be enable by adding /think to the system prompt and prefilling <think> or disabled by adding /nothink and prefilling <think>\n\n</think> instead.

Avoid using rep pen or dry if you can, as reasoning is degraded by penalizing repetition. ~.5 temp and minp .06 should be good defaults.

If you use Sillytavern make sure to disable prepending names.

How to Get Started with the Model

I recommend using koboldcpp to get started.

Training Details

The dataset used consists of R1 reasoning traces, system prompts scraped from github, and real world usage data from WildChat and ShareGPT with assistant responses regenerated with Qwen3 235B for STEM and coding data, and V303/24 for chat and multilingual data. Trained with Axolotl using the usual rank 64 rslora+ death stack of LoRA optimizations.