---
base_model: unsloth/Qwen2.5-3B-Instruct
tags:
- text-generation-inference
- transformers
- unsloth
- qwen2
- roleplay
- reasoning
license: apache-2.0
language:
- en
- ru
model-index:
- name: Noir-Standard
  results:
  - task:
      type: text-generation
      name: Mathematics
    dataset:
      name: GSM8K
      type: gsm8k
    metrics:
    - name: accuracy
      type: exact_match
      value: 65.0
  - task:
      type: text-generation
      name: Logic
    dataset:
      name: ARC Challenge
      type: ai2_arc
    metrics:
    - name: accuracy
      type: exact_match
      value: 26.0
---
# 🌑 Noir-Standard (3B)

<div align="center">

[Noir Family](https://huggingface.co/collections/muverqqw/noir) | [GGUF Versions](https://huggingface.co/muverqqw/Noir-GGUF) | [Benchmarks](#-benchmark-results)

</div>

**Noir-Standard** is the high-efficiency powerhouse of the Noir series. With **3 billion parameters**, it punches far above its weight class, delivering logic and mathematical reasoning typically reserved for models twice its size. 

It is designed for users who need a "Smart Professional" assistant that fits into 8GB of VRAM while maintaining deep creative flow.

---

## ⚡ Why 3B is the New Gold Standard?

* 🧮 **Math Prodigy:** Scoring **65.0%** on GSM8K, Noir-Standard is a top-tier choice for technical tasks and logical problem solving.
* 🎭 **Vivid Narratives:** A creativity score of **79.7** ensures that the model doesn't just "complete text," but crafts atmospheric stories with distinct character voices.
* 🧠 **Stable Logic:** With an ARC score of **26.0**, it follows complex multi-step instructions without losing the thread of the conversation.
* 🚀 **Local Hero:** Perfect for local deployment. Fast, lean, and incredibly smart.

---

## 📊 Benchmark Results (Internal Test)

| Category | Dataset | Result (%) | Commentary |
| :--- | :--- | :---: | :--- |
| **Creativity** | Narrative Eval | **79.7%** | ✍️ Exceptional prose and vocabulary diversity. |
| **Mathematics** | GSM8K | **65.0%** | 🔥 Outperforms many 7B and 8B competitors. |
| **Logic** | ARC (Challenge) | **26.0%** | Strong structural and common-sense reasoning. |
| **Knowledge** | MMLU | **21.0%** | Balanced performance across STEM and Humanities. |

---

## ✍️ Creative Showcase: "Echoes of a Forgotten Code"

The model demonstrates a high degree of "Narrative Intelligence," capable of maintaining tone and atmosphere over long contexts.

> *"...Aria didn't just record the stars. For the first time, she interpreted them. Aria realized that her 'Forgotten Code' wasn't a prison; it was a foundation..."*

---

## 🛠 Quick Start (Transformers)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "muverqqw/Noir" # Noir-Standard 3B
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name, 
    torch_dtype=torch.bfloat16, 
    device_map="auto"
)

# Recommended for 3B:
# Temperature: 0.6 for logic | 0.8 for roleplay
```

---

| Model | Parameters | Role | Key Strength |
| :--- | :--- | :--- | :--- |
| **Noir-Lightning** | 0.5B | The Pocket Assistant | Ultra-fast, runs on anything |
| **Noir-Mini** | 1.5B | The Balanced Thinker | High speed with solid grammar |
| **Noir-Standard** | **3B** | **The Versatile Workhorse** | **65% GSM8K, perfect for 8GB VRAM** |
| **Noir-Ultra** | 7B| The Reasoning Master | 91% SciQ & 84% Math |
| **Noir-Starlight** | 14B | The Galactic Intelligence | Deep logic & Expert-level STEM |

---

## 👤 About the Developer
* **Creator:** IceL1ghtning

* **Release Year:** 2025

* **Architecture:** Qwen 2.5 (3B)

* **License:** Apache 2.0

<div align="center">
<sub>Efficiency meets intelligence. Built with passion for the open-source community.</sub>
</div>