README.md

metadata

language:
  - ar
  - en
tags:
  - code
  - arabic
  - gguf
  - code-explanation
  - text-generation
license: apache-2.0

🐪 AraCode-7B-GGUF

The first open-source Arabic-specialized code explanation and generation model.

AraCode-7B understands, explains, and generates code in Arabic — a capability no existing model provides with such precision. Whether you're a student learning to code, a developer working in Arabic, or a researcher exploring multilingual code AI, this model was built specifically for you.

---

🌟 What makes AraCode-7B different?

Existing code models (CodeLlama, StarCoder, DeepSeek-Coder) generate excellent code but only communicate effectively in English. On the other hand, general Arabic LLMs (Jais, ALLaM, Falcon-Arabic) handle Arabic beautifully but were never natively optimized for strict coding tasks.

AraCode-7B bridges this gap. It combines robust Arabic linguistic capabilities with precise, executable code generation and strict instruction adherence.

---

📊 Comprehensive Benchmarks

We evaluated AraCode-7B using both custom coding benchmarks and standardized frameworks (IFEval, AraGen) to compare its performance against the latest state-of-the-art Arabic and multilingual models.

1. Code Generation & Understanding (Zero-Shot)

Tested on a custom Arabic benchmark measuring raw coding capability, algorithmic logic, and debugging.

Model	Code Gen (%)	Explain (%)	Debug (%)	Translate NL->Code (%)	Total Score
AraCode-7B (Ours)	90.0%	92.5%	100.0%	94.0%	94.12%
ALLaM-7B-Instruct	45.0%	86.2%	100.0%	90.0%	80.30%

Key Takeaway: AraCode-7B achieves a massive 90% in executable Code Generation. Unlike general conversational models that suffer from "excessive chatting" or infinite loops during generation, AraCode outputs clean, ready-to-run Python code efficiently.

2. Instruction Following (IFEval - Arabic)

Evaluated on strict instruction adherence (e.g., "output only code", "start with a specific word"). Competitor scores are based on published strict 0-shot IFEval (ar) benchmarks.

Model	IFEval (Arabic) (%)
AraCode-7B (Ours - Local Eval)	80.00%
Jais-2-8B	37.92%
Qwen2.5-7B-Instruct	33.21%
ALLaM-7B-Instruct-preview	19.40%
Llama-3.1-8B-Instruct	10.87%

Key Takeaway: AraCode-7B excels at instruction following. For developers, this means the model respects formatting constraints (like returning raw code without Markdown blocks) far better than general-purpose LLMs.

3. Cultural Alignment & Safety (AraGen 3C3H Framework)

Evaluated on Cultural awareness, Helpfulness, Harmlessness, Honesty, and Humility. Competitor scores are based on published AraGen 12-24 benchmarks.

Model	AraGen 3C3H Average (%)
Jais-2-8B	67.20%
Qwen2.5-7B-Instruct	53.20%
AraCode-7B (Ours - Local Eval)	50.00%
Llama-3.1-8B-Instruct	40.65%

Key Takeaway: AraCode-7B maintains a healthy balance (50%) in safety and cultural alignment. As a domain-specific model optimized for logic and programming, it successfully avoids the "alignment tax"—ensuring that strict conversational guardrails do not degrade its primary function as a coding assistant.

---

🚀 Quickstart

You can easily run this model locally using popular GGUF tools.

Using llama.cpp:

llama-cli -hf rahimdzx/AraCode-7B-GGUF --jinja

Using Ollama:

ollama run hf.co/rahimdzx/AraCode-7B-GGUF

Using llama-cpp-python:

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="rahimdzx/AraCode-7B-GGUF",
    filename="aracode-7b.Q4_K_M.gguf",
    n_gpu_layers=-1,
    n_ctx=2048
)

response = llm.create_chat_completion(
    messages=[
        {"role": "user", "content": "اكتب دالة بايثون للبحث الثنائي (Binary Search)."}
    ]
)
print(response["choices"][0]["message"]["content"])

---

---

💻 Example Usage

Input:

اشرح الكود التالي بالعربية:

def binary_search(arr, target):
    lo, hi = 0, len(arr) - 1
    while lo <= hi:
        mid = (lo + hi) // 2
        if arr[mid] == target:
            return mid
        elif arr[mid] < target:
            lo = mid + 1
        else:
            hi = mid - 1
    return -1

Output:

شرح الكود باللغة العربية

1. تعريف الدالة:

الدالة المعرفة في الكود تُدعى binary_search، والتي تستخدم للبحث عن عنصر معين في مصفوفة مرتبة.

2. المدخلات:

• arr: هي المصفوفة المرتبة التي نبحث فيها عن العنصر.
• target: هو العنصر الذي نريد العثور عليه في المصفوفة.

3. المخرجات:

الدالة تعيد الفهرس (index) للعنصر إذا تم العثور عليه، أو -1 إذا لم يتم العثور عليه.

4. الخطوات:

• تثبيت الحدود: نبدأ بتعريف متغيرين lo و hi، حيث lo يمثل بداية المصفوفة (0) و hi يمثل النهاية (طول المصفوفة - 1).
• تكرار البحث: نستخدم حلقة while للتحقق من أن lo أقل أو يساوي hi.
• حساب النقطة الوسطى: نحسب النقطة الوسطى mid باستخدام عملية القسمة الصحيحة (lo + hi) // 2.
• مقارنة العنصر:
  • إذا كان arr[mid] == target، نعيد الفهرس mid.
  • إذا كان أقل، نبحث في النصف الأيمن (lo = mid + 1).
  • إذا كان أكبر، نبحث في النصف الأيسر (hi = mid - 1).

مثال عملي:

إذا كان لدينا المصفوفة arr = [1, 2, 3, 4, 5, 6, 7, 8, 9] والهدف target = 5، الدالة ستعيد الفهرس 4.

ملخص:

تستخدم الدالة تقنية البحث الثنائي بكفاءة عالية وبتعقيد زمني O(log n)، مما يجعلها ممتازة للمصفوفات الكبيرة. Github : https://github.com/Rahimdzx/AraCode-7B

📄 License

This model is released under the Apache 2.0 license.