Violet-Magcap-12B: A Merged Reasoning Model
Violet-Magcap-12B is a 12 billion parameter language model developed by Nitral-AI, created through an unconventional merging process. It combines the strengths of two base models: inflatebot/MN-12B-Mag-Mell-R1 and Nitral-AI/Captain-Eris_Violet-GRPO-v0.420.
Key Capabilities & Architecture
- Unique Merge Architecture: The model's foundation is a merge of Mag-Mell-R1 and Captain-Eris, followed by a post-merge Supervised Fine-Tuning (SFT) phase.
- Enhanced Reasoning: It underwent a double SFT process using fresh reasoning data, specifically designed to boost its logical problem-solving abilities.
- Philosophical & Complex Text Generation: The model is noted for its capacity to handle complex philosophical concepts and generate nuanced text, often with an intriguing 'vibe check' that suggests it can make users "question their existence."
Usage & Formats
- Prompt Formats: Supports specific prompt formats including a "Reasoning Block + Prefix" and standard "ChatML Format," alongside "Quick Reply's" for diverse interaction styles.
- Quantized Versions: Available in optimized quantized versions, including GGUF (Lewdiculus Imatrix) and ExL2 (Nitral 4bpw), for efficient deployment.
- SillyTavern Presets: Provides dedicated presets for integration with SillyTavern, indicating suitability for role-playing and interactive narrative applications.
Ideal Use Cases
- Applications requiring advanced logical reasoning and problem-solving.
- Generating complex, philosophical, or thought-provoking content.
- Interactive storytelling and role-playing scenarios, especially with SillyTavern integration.