ArabFuncBench is the first natively constructed Arabic benchmark for evaluating function calling (tool use) in large language models. All utterances, tool descriptions, and argument values are written in Modern Standard Arabic (MSA) — not translated from English.

Paper: [ArabFuncBench: A Native Arabic Benchmark for Evaluating Function Calling in Large Language Models] (under review)
Authors: Lamyaa Sadouk — Ecole Marocaine des Sciences de l'Ingénieur, Casablanca, Morocco | Taoufiq Gadi — Laboratoire IRM, Morocco
Dataset: https://huggingface.co/datasets/lsadouk1111/ArabFuncBench
Code: https://github.com/lsadouk/ArabFuncBench
License: CC-BY 4.0

Dataset Summary

ArabFuncBench comprises 1,000 examples across five real-world Arabic service domains:

Domain	Tools	Positive	Negative	Total
Education	10	160	40	200
E-commerce	10	160	40	200
Healthcare	10	160	40	200
Islamic Services	10	160	40	200
Government	10	160	40	200
Total	50	800	200	1,000

Motivation

Arabic-speaking organizations increasingly deploy AI-powered digital services — government portals, hospital systems, banking interfaces, e-commerce platforms. Function calling is critical for these systems: the LLM must translate Arabic user intent into structured, executable API calls with correctly extracted Arabic argument values.

Despite rapid growth in Arabic LLM development, no standardized benchmark existed for evaluating this capability. ArabFuncBench fills this gap.

Dataset Structure

Files

arab_func_bench_examples.json — 1,000 evaluation examples
arab_func_bench_tools.json — 50 tool definitions in OpenAI JSON schema format
all_metrics.json — Evaluation results for all 7 models

Example Format

{
  "id": "islamic_services_calculate_prayer_times_001",
  "domain": "islamic_services",
  "utterance": "متى موعد صلاة الفجر في الرياض اليوم؟",
  "is_negative": false,
  "expected_function": "calculate_prayer_times",
  "expected_arguments": {
    "city": "الرياض",
    "date": "اليوم"
  },
  "available_tools": [
    "calculate_prayer_times",
    "get_hadith",
    "find_nearest_mosque"
  ]
}

Tool Definition Format

{
  "name": "calculate_prayer_times",
  "description": "يحسب أوقات الصلاة الخمس لمدينة معينة وتاريخ محدد",
  "parameters": {
    "type": "object",
    "properties": {
      "city": {
        "type": "string",
        "description": "اسم المدينة"
      },
      "date": {
        "type": "string",
        "description": "التاريخ المطلوب"
      }
    },
    "required": ["city", "date"]
  }
}

Evaluation Metrics

Three metrics are defined for Arabic function calling evaluation:

Tool Selection Accuracy (TSA): Whether the model correctly identifies the function to call. For negative examples, TSA=1 if the model correctly returns null.

Argument Extraction F1 (AEF1): Character-level fuzzy matching (θ=0.65) between predicted and expected argument values, conditioned on correct tool selection.

Language Compliance Rate (LCR): Whether argument values are returned in Arabic rather than English. Numeric values (IDs, dates) are excluded from this check.

Benchmark Results

Model	Type	Size	TSA	AEF1	LCR
Llama-3.3-70B	Multilingual	70B	0.997	0.910	0.949
Qwen2.5-7B	Multilingual	7B	0.992	0.827	0.942
GPT-4o-mini	Multilingual	Proprietary	0.978	0.876	0.932
Claude Haiku	Multilingual	Proprietary	0.924	0.927	0.938
ALLaM-7B	Arabic specialized	7B	0.923	0.830	0.914
Fanar-9B	Arabic specialized	9B	0.871	0.861	0.881
AceGPT-7B	Arabic specialized	7B	0.224	0.708	0.833

Key Findings

Instruction-tuned multilingual models consistently outperform Arabic-specialized models
ALLaM-7B nearly matches Claude Haiku on TSA (0.923 vs 0.924) — instruction tuning quality matters more than Arabic specialization
AceGPT-7B defaults to null on 97% of positive examples — Arabic fine-tuning without structured output training is insufficient
Education is the hardest domain (avg TSA 0.929) due to semantic overlap between functionally adjacent tools
Fanar-9B and ALLaM-7B show highest Type 4 error rates — Arabic-specialized models revert to English argument values under JSON output constraints

Domains

Education: School scheduling, student grades, attendance, homework, transcripts

E-commerce: Product search, order tracking, payments, discounts, returns

Healthcare: Appointments, lab results, medications, prescriptions, health reminders

Islamic Services: Prayer times, Quran verses, Hijri calendar, Zakat calculation, Qibla direction

Government: Passport renewal, vehicle registration, scholarships, driving license, birth certificates

Construction

Tool definitions: 50 manually reviewed tools in OpenAI JSON schema format
Example generation: Claude Haiku API with domain-specific Arabic prompts
Quality control: 200 examples manually validated (20% of dataset)
Systematic fixes: Arabic-Indic numeral normalization, year reference updates, Latin coupon code repositioning
Negative examples: Fully regenerated after detecting 20% label error rate in initial generation

Evaluation Protocol

Zero-shot evaluation — no fine-tuning, no task-specific adaptation. Each model receives the Arabic utterance and available tool definitions in JSON schema format and must return a structured JSON response.

Citation

@misc{sadouk2026arabfuncbench, title={ArabFuncBench: A Native Arabic Benchmark for Evaluating Function Calling in Large Language Models}, author={Sadouk, Lamyaa and Gadi, Taoufiq}, year={2026}, howpublished={ResearchGate preprint}, note={Under review at ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP).} }

License

This dataset is released under CC-BY 4.0. You are free to use, share, and adapt it for any purpose, provided appropriate credit is given.

Contact

Lamyaa Sadouk — Ecole Marocaine des Sciences de l'Ingénieur, Casablanca, Morocco

Downloads last month: 16

Total file size:

622 kB