ewald1976/Red-Rosie-12B
Red Rosie 12B is an experimental SLERP merge model created by ewald1976, combining TheDrummer/Rocinante-X-12B-v1 and WokeAI/Tankie-DPE-12B-SFT-v2. This 12 billion parameter model integrates Rocinante-X's top-tier roleplay capabilities with a subtle influence from Tankie-DPE's Marxist-Leninist ideological training. It is designed for personal research to explore the effects of merging models with contrasting characteristics, particularly in how ideological biases manifest.
Loading preview...
Red Rosie 12B: An Experimental SLERP Merge
Red Rosie 12B is an experimental 12 billion parameter language model developed by ewald1976, created using a SLERP (Spherical Linear Interpolation) merge via mergekit. This unique model combines two distinct 12B base models:
- TheDrummer/Rocinante-X-12B-v1: Known for its top-tier roleplay capabilities within the Nemo family.
- WokeAI/Tankie-DPE-12B-SFT-v2: A model post-post-trained towards Marxist-Leninist ideology.
Merge Method and Characteristics
The merge strategically applies Tankie-DPE's influence to the middle layers (12–28) at a weight of 0.20, while the early (0–12) and late (28–40) layers remain pure Rocinante-X. This method aims to preserve the embedding coherence and output quality of Rocinante-X while allowing a nuanced ideological character from Tankie-DPE to emerge in deeper reasoning.
Warnings and Intended Use
This model is explicitly experimental and its behavior may be unpredictable. Due to the influence of Tankie-DPE, it carries an ideological bias towards Marxist-Leninist positions, which may surface even with neutral system prompts. Red Rosie 12B is not intended for production use and is shared solely for personal research purposes to investigate the outcomes of merging models with divergent training objectives.