Sakura-Sniper-12B/README.md

---
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
---

<img src="https://i.postimg.cc/43sf2CxF/Gemini-Generated-Image-oknnsxoknnsxoknn.png" alt="cover" border="0" width="1024px">


# 🌸 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).