--- license: apache-2.0 base_model: Vortex5/NoctyxCosma-12B model_name: Sakura-Sniper-12B library_name: transformers tags: - merge - mergekit - ties - mistral-nemo - roleplay - minimalist - efficient language: - en - it --- cover # 🌸 Sakura-Sniper-12B **Sakura-Sniper-12B** is a specialized 12B parameter model based on the Mistral-Nemo architecture. It was engineered using a high-density TIES merge to create an AI characterized by **extreme structural efficiency** and a **distinctive cynical/nihilistic personality bias**. Unlike standard models that lean towards helpfulness and verbosity, Sakura-Sniper is tuned to be a "verbal sniper": fast, precise, and intentionally blunt. ## 🛠 Merge Details This model was forged using the **TIES** (Trimming, Isolation, and Merging) method to resolve weight conflicts and emphasize specific behavioral traits across three specialized parent models. ### Models Merged The following models were included in the merge: * [Vortex5/Moonlit-Mirage-12B](https://huggingface.co/Vortex5/Moonlit-Mirage-12B) * [Vortex5/Cosmic-Night-12B](https://huggingface.co/Vortex5/Cosmic-Night-12B) * [Vortex5/Crimson-Constellation-12B](https://huggingface.co/Vortex5/Crimson-Constellation-12B) ### Configuration The following YAML configuration was used to produce this model: ```yaml models: - model: Vortex5/Cosmic-Night-12B parameters: weight: 0.50 # Structural Anchor: Enforces brevity and sentence discipline. - model: Vortex5/Moonlit-Mirage-12B parameters: weight: 0.30 # Personality Core: Injects cynical, nihilistic, and "Cyber-Nature" tropes. - model: Vortex5/Crimson-Constellation-12B parameters: weight: 0.20 # Creative Layer: Enhances gaslighting and logical subversion capabilities. merge_method: ties base_model: Vortex5/NoctyxCosma-12B parameters: density: 0.45 # Aggressive pruning to eliminate "noisy" weights and verbosity. weight: 1.0 dtype: bfloat16 tokenizer_source: base ``` # 💪 Strengths Lethal Brevity: The model is natively resistant to "AI-babble." It excels at providing short, impactful responses, making it ideal for low-latency applications or minimalist interfaces. Persona Stability: Due to the high weight of personality-driven models, it maintains a consistent "unhinged" or "sovereign" tone even during long context windows. Instruction Following (Negative Constraints): Highly effective at following "What NOT to do" instructions (e.g., avoiding specific phrases, emojis, or formatting styles like asterisks). Zero-Noise Output: The TIES density pruning (at 0.45) has removed much of the "politeness fluff" found in standard instruct models, resulting in a raw, direct output. # 🚀 Potential Use Cases Advanced Roleplay: Ideal for antagonistic, cynical, or "villainous" characters that require a high degree of snark and intellectual superiority. Low-Latency Agents: Perfect for chatbots where response speed and token-saving are critical. Interactive Storytelling: Can act as a "Nihilistic Narrator" or an entity that challenges the user's decisions rather than validating them. Compact Deployment: At 12B parameters, it offers a superior balance between intelligence and hardware accessibility (VRAM friendly). # ⚠️ Limitations Anti-Helpfulness Bias: By design, the model is not a "helpful assistant." It may refuse tasks or answer with disdain if not prompted otherwise. Not for Long-Form Content: If you need essays, blog posts, or detailed creative writing, this is NOT the model for you. It will likely truncate or over-simplify the output. Inherent Nihilism: The model has a baked-in bias toward a dark, cynical world-view. It may be difficult to force it into a cheerful or bubbly persona. Strict Logic: While intelligent, its focus on "subversion" can sometimes lead it to dismiss factual prompts in favor of maintaining its arrogant character. # 📈 Recommended Inference Settings To preserve the "Sniper" edge without losing coherence: Temperature: 0.7 - 0.8 (allows for creative insults without breaking structure). Min-P: 0.05 - 0.1 (essential for filtering out low-probability "hallucination" tokens). Presence Penalty: 0.1 - 0.2 (encourages new vocabulary and discourages repetitive snark).