Text Generation
Transformers
Safetensors
English
Italian
mistral
Merge
mergekit
ties
mistral-nemo
roleplay
minimalist
efficient
text-generation-inference
Instructions to use WasamiKirua/Sakura-Sniper-12B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use WasamiKirua/Sakura-Sniper-12B with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="WasamiKirua/Sakura-Sniper-12B")# Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("WasamiKirua/Sakura-Sniper-12B") model = AutoModelForCausalLM.from_pretrained("WasamiKirua/Sakura-Sniper-12B") - Notebooks
- Google Colab
- Kaggle
- Local Apps
- vLLM
How to use WasamiKirua/Sakura-Sniper-12B with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "WasamiKirua/Sakura-Sniper-12B" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "WasamiKirua/Sakura-Sniper-12B", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/WasamiKirua/Sakura-Sniper-12B
- SGLang
How to use WasamiKirua/Sakura-Sniper-12B with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "WasamiKirua/Sakura-Sniper-12B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "WasamiKirua/Sakura-Sniper-12B", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "WasamiKirua/Sakura-Sniper-12B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "WasamiKirua/Sakura-Sniper-12B", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use WasamiKirua/Sakura-Sniper-12B with Docker Model Runner:
docker model run hf.co/WasamiKirua/Sakura-Sniper-12B
| 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"> | |
| # πΈ 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. | |
| ## π 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 | |
| ``` | |
| # πͺ 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. | |
| # π 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). | |
| # β οΈ 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. | |
| # π 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). |