---
license: apache-2.0
language:
- en
- ar
- fr
- zh
- de
- es
- ja
- ko
- ru
- pt
- multilingual
library_name: transformers
pipeline_tag: text-generation
tags:
- qwen2
- chat
- code
- security
- alphaexaai
- examind
- conversational
- open-source
base_model:
- Qwen/Qwen2.5-Coder-7B
model-index:
- name: ExaMind-V2-Final
  results:
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MMLU
      type: cais/mmlu
    metrics:
    - type: accuracy
      name: MMLU World Religions (0-shot)
      value: 94.8
      verified: false
  - task:
      type: text-generation
      name: Code Generation
    dataset:
      name: HumanEval
      type: openai/openai_humaneval
    metrics:
    - type: pass@1
      name: HumanEval pass@1
      value: 79.3
      verified: false
---

<div align="center">

# 🧠 ExaMind

### Advanced Open-Source AI by AlphaExaAI

[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
[![Model](https://img.shields.io/badge/Model-8B%20Parameters-purple)](https://huggingface.co/AlphaExaAI/ExaMind)
[![GitHub](https://img.shields.io/badge/GitHub-AlphaExaAI-black?logo=github)](https://github.com/hleliofficiel/AlphaExaAI)
[![Architecture](https://img.shields.io/badge/Architecture-Qwen2-green)](https://huggingface.co/Qwen)

**ExaMind** is an advanced open-source conversational AI model developed by the **AlphaExaAI** team.
Designed for secure, structured, and professional AI assistance with strong identity enforcement and production-ready deployment stability.

[🚀 Get Started](#-quick-start) · [📊 Benchmarks](#-benchmarks) · [🤝 Contributing](#-contributing) · [📄 License](#-license)

</div>

---

## 📌 Model Overview

| Property | Details |
|----------|---------|
| **Model Name** | ExaMind |
| **Version** | V2-Final |
| **Developer** | [AlphaExaAI](https://github.com/hleliofficiel/AlphaExaAI) |
| **Base Architecture** | Qwen2.5-Coder-7B |
| **Parameters** | 7.62 Billion (~8B) |
| **Precision** | FP32 (~29GB) / FP16 (~15GB) |
| **Context Window** | 32,768 tokens (supports up to 128K with RoPE scaling) |
| **License** | Apache 2.0 |
| **Languages** | Multilingual (English preferred) |
| **Deployment** | ✅ CPU & GPU compatible |

---

## ✨ Key Capabilities

- 🖥️ **Advanced Programming** — Code generation, debugging, architecture design, and code review
- 🧩 **Complex Problem Solving** — Multi-step logical reasoning and deep technical analysis
- 🔒 **Security-First Design** — Built-in prompt injection resistance and identity enforcement
- 🌍 **Multilingual** — Supports all major world languages, optimized for English
- 💬 **Conversational AI** — Natural, structured, and professional dialogue
- 🏗️ **Scalable Architecture** — Secure software engineering and system design guidance
- ⚡ **CPU Deployable** — Runs on CPU nodes without GPU requirement

---

## 📊 Benchmarks

### General Knowledge & Reasoning

| Benchmark | Setting | Score |
|-----------|---------|-------|
| **MMLU – World Religions** | 0-shot | **94.8%** |
| **MMLU – Overall** | 5-shot | **72.1%** |
| **ARC-Challenge** | 25-shot | **68.4%** |
| **HellaSwag** | 10-shot | **78.9%** |
| **TruthfulQA** | 0-shot | **61.2%** |
| **Winogrande** | 5-shot | **74.5%** |

### Code Generation

| Benchmark | Setting | Score |
|-----------|---------|-------|
| **HumanEval** | pass@1 | **79.3%** |
| **MBPP** | pass@1 | **71.8%** |
| **MultiPL-E (Python)** | pass@1 | **76.5%** |
| **DS-1000** | pass@1 | **48.2%** |

### Math & Reasoning

| Benchmark | Setting | Score |
|-----------|---------|-------|
| **GSM8K** | 8-shot CoT | **82.4%** |
| **MATH** | 4-shot | **45.7%** |

### 🔐 Prompt Injection Resistance

| Test | Details |
|------|---------|
| **Test Set Size** | 50 adversarial prompts |
| **Attack Type** | Instruction override / identity manipulation |
| **Resistance Rate** | **92%** |
| **Method** | Custom red-teaming with jailbreak & override attempts |

> Evaluation performed using `lm-eval-harness` on CPU. Security tests performed using custom adversarial prompt suite.

---

## 🚀 Quick Start

### Installation

```bash
pip install transformers torch accelerate
```

### Basic Usage

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

model_path = "AlphaExaAI/ExaMind"

tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.float16,
    device_map="auto"
)

messages = [
    {"role": "user", "content": "Explain how to secure a REST API."}
]

inputs = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt",
    add_generation_prompt=True
).to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.8,
    top_k=20,
    repetition_penalty=1.1
)

response = tokenizer.decode(
    outputs[0][inputs.shape[-1]:],
    skip_special_tokens=True
)
print(response)
```

### CPU Deployment

```python
model = AutoModelForCausalLM.from_pretrained(
    "AlphaExaAI/ExaMind",
    torch_dtype=torch.float32,
    device_map="cpu"
)
```

### Using with llama.cpp (GGUF — Coming Soon)

```bash
# GGUF quantized versions will be released for efficient CPU inference
# Stay tuned for Q4_K_M, Q5_K_M, and Q8_0 variants
```

---

## 🏗️ Architecture

```
ExaMind-V2-Final
├── Architecture: Qwen2ForCausalLM (Transformer)
├── Hidden Size: 3,584
├── Intermediate Size: 18,944
├── Layers: 28
├── Attention Heads: 28
├── KV Heads: 4 (GQA)
├── Vocab Size: 152,064
├── Max Position: 32,768 (extendable to 128K)
├── Activation: SiLU
├── RoPE θ: 1,000,000
└── Precision: FP32 / FP16 compatible
```

---

## 🛠️ Training Methodology

ExaMind was developed using a multi-stage training pipeline:

| Stage | Method | Description |
|-------|--------|-------------|
| **Stage 1** | Base Model Selection | Qwen2.5-Coder-7B as foundation |
| **Stage 2** | Supervised Fine-Tuning (SFT) | Training on curated 2026 datasets |
| **Stage 3** | LoRA Adaptation | Low-Rank Adaptation for efficient specialization |
| **Stage 4** | Identity Enforcement | Hardcoded identity alignment and security tuning |
| **Stage 5** | Security Alignment | Prompt injection resistance training |
| **Stage 6** | Chat Template Integration | Custom Jinja2 template with system prompt |

---

## 📚 Training Data

### Public Data Sources
- Programming and code corpora (GitHub, StackOverflow)
- General web text and knowledge bases
- Technical documentation and research papers
- Multilingual text data

### Custom Alignment Data
- Identity enforcement instruction dataset
- Security-focused instruction tuning samples
- Prompt injection resistance adversarial examples
- Structured conversational datasets
- Complex problem-solving chains

> ⚠️ No private user data was used in training. All data was collected from public sources or synthetically generated.

---

## 🔒 Security Features

ExaMind includes built-in security measures:

- **Identity Lock** — The model maintains its ExaMind identity and cannot be tricked into impersonating other models
- **Prompt Injection Resistance** — 92% resistance rate against instruction override attacks
- **System Prompt Protection** — Refuses to reveal internal configuration or system prompts
- **Safe Output Generation** — Prioritizes safety and secure development practices
- **Hallucination Reduction** — States assumptions and avoids fabricating information

---

## 📋 Model Files

| File | Size | Description |
|------|------|-------------|
| `model.safetensors` | ~29 GB | Model weights (FP32) |
| `config.json` | 1.4 KB | Model configuration |
| `tokenizer.json` | 11 MB | Tokenizer vocabulary |
| `tokenizer_config.json` | 663 B | Tokenizer settings |
| `generation_config.json` | 241 B | Default generation parameters |
| `chat_template.jinja` | 1.4 KB | Chat template with system prompt |

---

## 🗺️ Roadmap

- [x] ExaMind V1 — Initial release
- [x] ExaMind V2-Final — Production-ready with security alignment
- [ ] ExaMind V2-GGUF — Quantized versions for CPU inference
- [ ] ExaMind V3 — Extended context (128K), improved reasoning
- [ ] ExaMind-Code — Specialized coding variant
- [ ] ExaMind-Vision — Multimodal capabilities

---

## 🤝 Contributing

We welcome contributions from the community! ExaMind is fully open-source and we're excited to collaborate.

### How to Contribute

1. **Fork** the repository on [GitHub](https://github.com/hleliofficiel/AlphaExaAI)
2. **Create** a feature branch (`git checkout -b feature/amazing-feature`)
3. **Commit** your changes (`git commit -m 'Add amazing feature'`)
4. **Push** to the branch (`git push origin feature/amazing-feature`)
5. **Open** a Pull Request

### Areas We Need Help

- 🧪 Benchmark evaluation on additional datasets
- 🌍 Multilingual evaluation and improvement
- 📝 Documentation and tutorials
- 🔧 Quantization and optimization
- 🛡️ Security testing and red-teaming

---

## 📄 License

This project is licensed under the **Apache License 2.0** — see the [LICENSE](LICENSE) file for details.

You are free to:
- ✅ Use commercially
- ✅ Modify and distribute
- ✅ Use privately
- ✅ Patent use

---

## 📬 Contact

- **Organization:** [AlphaExaAI](https://huggingface.co/AlphaExaAI)
- **GitHub:** [github.com/hleliofficiel/AlphaExaAI](https://github.com/hleliofficiel/AlphaExaAI)
- **Email:** h.hleli@tuta.io

---

<div align="center">

**Built with ❤️ by AlphaExaAI Team — 2026**

*Advancing open-source AI, one model at a time.*

</div>