Arabic Function Calling Dataset (50K+ Samples)

Dataset Description

This is the first comprehensive Arabic function calling dataset designed for training and evaluating LLMs on Arabic tool use capabilities. The dataset covers:

• 5 Arabic Dialects: MSA (Modern Standard Arabic), Egyptian, Gulf, Levantine, and Maghrebi
• 8 Real-World Domains: Islamic services, government services, banking/finance, e-commerce, travel, healthcare, weather, and utilities
• 36+ Functions: Covering practical Arabic-world use cases

Key Features

• 50,810 total samples (45,734 positive + 5,076 negative examples)
• Multi-dialect coverage reflecting real Arabic usage patterns
• Culturally authentic - designed for Arab world use cases, not translated from English
• GPT-4o generated with quality validation
• OpenAI-compatible format for easy fine-tuning

Dataset Statistics

By Dialect

By Domain

Sample Types

Dataset Structure

Each sample contains:

{
  "id": "gen_00001",
  "query_ar": "عايز أعرف مواعيد الصلاة في القاهرة النهارده",
  "query_en": "I want to know the prayer times in Cairo today",
  "function_name": "get_prayer_times",
  "arguments": {"city": "القاهرة", "date": "today"},
  "dialect": "Egyptian",
  "domain": "islamic_services",
  "requires_function": true
}

Fields Description

Available Functions

Islamic Services (الخدمات الإسلامية)

• get_prayer_times - مواقيت الصلاة
• calculate_zakat - حساب الزكاة
• get_qibla_direction - اتجاه القبلة
• search_quran - البحث في القرآن
• get_hadith - البحث في الأحاديث
• calculate_inheritance - حساب المواريث

Government Services (الخدمات الحكومية)

• check_visa_status - حالة التأشيرة
• check_iqama_status - صلاحية الإقامة
• check_traffic_violations - المخالفات المرورية
• book_government_appointment - حجز موعد حكومي
• calculate_end_of_service - مكافأة نهاية الخدمة

Banking & Finance (البنوك والمالية)

• convert_currency - تحويل العملات
• get_gold_price - سعر الذهب
• calculate_loan - حساب القرض
• transfer_money - تحويل الأموال

E-commerce (التجارة الإلكترونية)

• track_shipment - تتبع الشحنة
• compare_prices - مقارنة الأسعار
• calculate_customs - حساب الجمارك
• order_food - طلب طعام

Travel (السفر)

• search_flights - البحث عن رحلات
• search_hotels - البحث عن فنادق
• search_umrah_packages - باقات العمرة

Healthcare (الصحة)

• book_doctor_appointment - حجز موعد طبي
• search_medications - البحث عن الأدوية
• check_insurance_coverage - تغطية التأمين

Weather (الطقس)

• get_weather - حالة الطقس
• get_air_quality - جودة الهواء

Utilities (الأدوات)

• translate_text - الترجمة
• calculate_math - العمليات الحسابية
• set_reminder - إنشاء تذكير
• get_current_time - الوقت الحالي

Usage

Loading with Datasets Library

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("HeshamHaroon/Arabic_Function_Calling")

# Access splits
train_data = dataset["train"]
test_data = dataset["test"]

# Example usage
for sample in train_data.select(range(5)):
    print(f"Query: {sample['query_ar']}")
    print(f"Function: {sample['function_name']}")
    print(f"Dialect: {sample['dialect']}")
    print("---")

Fine-tuning Example

from transformers import AutoTokenizer, AutoModelForCausalLM
from datasets import load_dataset

dataset = load_dataset("HeshamHaroon/Arabic_Function_Calling")

def format_for_training(example):
    system = "أنت مساعد ذكي يفهم اللغة العربية. عند تلقي طلب، قم باستدعاء الدالة المناسبة."

    if example['requires_function']:
        response = f'{{"name": "{example["function_name"]}", "arguments": {example["arguments"]}}}'
    else:
        response = "لا حاجة لاستدعاء دالة"

    return {
        "messages": [
            {"role": "system", "content": system},
            {"role": "user", "content": example['query_ar']},
            {"role": "assistant", "content": response}
        ]
    }

formatted_dataset = dataset.map(format_for_training)

Evaluation Results

Validated on GPT-4o with function calling:

By Dialect Performance

Dataset Creation

Methodology

This dataset was created using a hybrid approach combining:

1. Seed Generation: 500 hand-crafted samples across all domains and dialects
2. Self-Instruct: GPT-4o expansion of seeds to 5,000 samples
3. Dialect Multiplication: Expanding across 5 Arabic dialects
4. Quality Filtering: Validation and deduplication

Why Native Arabic (Not Translation)?

Based on QCRI's research on Arabic Tool Calling:

• Pure translation introduces "Anglocentric biases"
• Bilingual (EN+AR) training achieves 0.99 recall vs 0.69 for single-language
• Native Arabic datasets capture cultural context better

Generation Details

• Model: Azure OpenAI GPT-4o
• Concurrency: 20 parallel requests with exponential backoff
• Temperature: 0.9 for diversity
• Validation: GPT-4o function calling with 36 tools

Limitations and Biases

Known Limitations

1. Synthetic Data: Generated by GPT-4o, may have model biases
2. Dialect Imbalance: Maghrebi dialect underrepresented (4.1%)
3. Domain Coverage: Limited to 8 domains, doesn't cover all use cases
4. Argument Values: Some generated values may be fictional

Recommended Use Cases

• Fine-tuning Arabic LLMs for function calling
• Benchmarking Arabic tool use capabilities
• Research on multilingual function calling
• Developing Arabic AI assistants

Not Recommended For

• Direct use in production without validation
• Legal or medical advice applications
• Financial decision-making systems

Citation

If you use this dataset, please cite:

@dataset{arabic_function_calling_2024,
  title={Arabic Function Calling Dataset},
  author={Hesham Haroon},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/HeshamHaroon/Arabic_Function_Calling}
}

Related Resources

• QCRI Arabic Tool Calling Research
• Trelis Function Calling Dataset
• xLAM Function Calling Dataset

License

This dataset is released under the Apache 2.0 License.

Contact

• Author: Hesham Haroon
• HuggingFace: @HeshamHaroon

---