---
base_model: unsloth/Qwen2.5-7B-Instruct # Уточни, если база другая
tags:
- text-generation-inference
- transformers
- unsloth
- qwen2
- reasoning
- science
- logic
- trl
- sft
license: apache-2.0
language:
- en
- ru
- zh
model-index:
- name: Noir-Ultra
  results:
  - task:
      type: text-generation
      name: STEM (Science)
    dataset:
      name: SciQ
      type: sciq
    metrics:
    - name: accuracy
      type: exact_match
      value: 91.0
  - task:
      type: text-generation
      name: Mathematics
    dataset:
      name: GSM8K
      type: gsm8k
    metrics:
    - name: accuracy
      type: exact_match
      value: 84.0
  - task:
      type: text-generation
      name: Logic
    dataset:
      name: ARC Challenge
      type: ai2_arc
    metrics:
    - name: accuracy
      type: exact_match
      value: 86.0
---

# 🌌 Noir-Ultra (7B)

<div align="center">

[Noir Family](https://huggingface.co/collections/muverqqw/noir) | [GGUF Versions](https://huggingface.co/muverqqw/Noir-GGUF) | [Technical Report](#-noir-ultra-evaluation-dashboard)

</div>

**Noir-Ultra** is the elite 7-billion parameter model of the Noir series. It represents a breakthrough in training efficiency: where previous 7B iterations required 6 epochs to reach stability, **Noir-Ultra achieved superior results in just a single epoch**.

This model is a "compact titan," delivering scientific reasoning and mathematical accuracy that rival much larger architectures.

---

## 🚀 The Ultra Advantage

* 🧬 **Unrivaled STEM:** With a **91.0% score on SciQ**, it is a specialized tool for scientific inquiry.
* 📐 **Mathematical Precision:** Scoring **84.0% on GSM8K**, it handles complex chains of thought with ease.
* 🧠 **Logical Depth:** An ARC-Challenge score of **86.0%** places it at the top of its weight class for reasoning tasks.

---

## 📊 Evaluation Dashboard (Noir Ultra Report)

Based on the latest evaluation, Noir-Ultra shows an exceptionally strong profile in technical domains:

| Domain | Benchmark | Result (%) | Status |
| :--- | :--- | :---: | :--- |
| **STEM** | SciQ | **91.0%** | 🏆 Master |
| **Logic** | ARC-C | **86.0%** | 🔥 Elite |
| **Math** | GSM8K | **84.0%** | ✅ Advanced |
| **Medicine** | MedQA | **65.0%** | 🩺 Competent |
| **Physics** | MMLU-Physics| **70.0%** | 🧪 Specialist |



---

## 📦 Noir Model Family Matrix

| 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 |

---

## 🛠 Quick Start (Transformers)

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

model_name = "muverqqw/Noir-Ultra" 
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name, 
    load_in_4bit=True, # Recommended for 8GB VRAM
    device_map="auto"
)
```

---

### 👤 Credits
* **Developer:** IceL1ghtning

* **Architecture:** Qwen 2.5 (7B)

* **Release:** 2026

* **License:** Apache 2.0

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