StarDust-12b-v2: A Specialized Role-Playing and Creative Writing Model

StarDust-12b-v2 is a 12 billion parameter language model, developed by Luni, that stands out for its exceptional capabilities in creative story writing and long-form role-playing. Merged using the DARE TIES method, it leverages several base models including Sao10K/MN-12B-Lyra-v3, nbeerbower/mistral-nemo-bophades-12B, anthracite-org/magnum-v2-12b, and Gryphe/Pantheon-RP-1.6-12b-Nemo.

Key Capabilities

• Vibrant and Creative Prose: Generates engaging and less generic narratives, capable of adapting tone from gentle to harsh as requested.
• Optimized for Role-playing: Specifically designed for extended role-play sessions, providing a stable and creative experience.
• Character Impersonation: Can be effectively used as an "Agent" by forcing it into character impersonation, though narratives may occasionally interject.
• Non-KTO Training: Utilizes a non-KTO magnum, which users report results in less "claudeism" and repetitive storytelling compared to its v1 counterpart.

Intended Use Cases

• Creative Story Writing: Ideal for generating imaginative stories and prose.
• Long-form Role-playing: Excels in interactive and extended role-play scenarios.
• Character-driven Narratives: Suitable for applications requiring the model to adopt and maintain specific character personas.

Limitations

• Not for General Conversation: The model is explicitly not recommended for direct conversational output, general purpose tasks, or direct instruction following.

Prompting

• ChatML Recommended: ChatML has been identified as the most effective prompting format for this model, though Mistral format is also supported.