Mosescreates/arabic-agent-eval

Arabic Agent Eval — Dataset Card

An open, installable Arabic function-calling benchmark with dialect splits.

Dataset summary

51 evaluation items spanning 6 categories and 5 dialects of Arabic, testing whether large language models can (a) select the right tool, (b) extract arguments from natural Arabic instructions, (c) preserve Arabic text in tool arguments instead of transliterating, and (d) understand dialectal framing.

Supported tasks

• Function-calling evaluation — items pair Arabic instructions with expected structured tool calls.
• Tool-use argument extraction — items declare expected_calls[i].arguments that the model's output is compared against via canonical structured-call grading.
• Dialect-robustness evaluation — dialect-labeled items enable cross-dialect variance analysis for any agent.

Languages

Arabic only. Five dialect registers: Modern Standard Arabic (msa), Gulf (gulf), Egyptian (egyptian), Levantine (levantine), Maghrebi (maghrebi).

Dataset structure

Each row is a single evaluation item. See docs/schema.md for full schema. Example:

{
  "id": "dialect_001",
  "category": "dialect_handling",
  "instruction": "ابي أحجز فندق في دبي بكرة",
  "dialect": "gulf",
  "available_functions": ["book_hotel", "search_flights", "get_weather"],
  "expected_calls": [
    {"function": "book_hotel", "arguments": {"city": "دبي", "check_in": "غداً"}}
  ],
  "difficulty": "easy"
}

Splits

Split	Items	Purpose
msa	32	Modern Standard Arabic baseline
gulf	10	Gulf dialect (Saudi, UAE, Kuwait, Qatar, Bahrain, Oman)
egyptian	3	Egyptian dialect
levantine	4	Levantine dialect (Jordan, Palestine, Syria, Lebanon)
maghrebi	2	Maghrebi dialect (Morocco, Algeria, Tunisia)
all	51	concatenation of the five dialect splits

Non-MSA dialects are undersampled in v0.1.0 (see docs/baselines.md). Use per-dialect averages as directional signal, not as final rankings.

Categories

Category	Arabic	Items	Weight
Simple Function Calling	استدعاء بسيط	10	0.20
Parameter Extraction	استخراج المعاملات	9	0.20
Multi-Step Reasoning	تفكير متعدد الخطوات	7	0.20
Dialect Handling	معالجة اللهجات	12	0.15
Tool Selection	اختيار الأداة	7	0.15
Error Recovery	معالجة الأخطاء	6	0.10

Tool registry

22 Arabic-context tool definitions ship alongside the items, covering Islamic services (prayer times, zakat, Quran search), GCC financial services (Tadawul stocks, multi-currency conversion, money transfer), regional consumer actions (ride-hailing, food delivery, hotel booking), and communication (WhatsApp, SMS, email). See data/functions.json.

Source data

Items are hand-written native Arabic instructions drawn from real regional use cases (Saudi, UAE, Egyptian, Levantine, Moroccan). Authored by the repository maintainer (a native Arabic speaker and digital-marketing practitioner). No data scraped or model-generated.

Personal and sensitive information

None. Items use fictitious names (e.g. أحمد, سارة, محمد) and generic city references.

Licensing

Data: CC-BY-4.0 (data/, dataset_card.md, docs/) Code: Apache-2.0 (arabic_agent_eval/, tests/, scripts/)

See LICENSES.md.

Citation

@software{arabic_agent_eval_2026,
  title = {Arabic Agent Eval: An open, dialect-split Arabic function-calling benchmark},
  author = {Abumazin, Mousa},
  year = {2026},
  url = {https://github.com/Moshe-ship/arabic-agent-eval}
}

Limitations and biases

• Dialect coverage is uneven in v0.1.0 (32 MSA / 10 Gulf / 4 Levantine / 3 Egyptian / 2 Maghrebi). Per-dialect statistics from the rare splits are directional only.
• GCC bias in the tool registry — Gulf-region contexts dominate (Saudi addresses, Tadawul, AED/SAR currency defaults).
• No intra-dialect paraphrase variance in v0.1.0. Scheduled for expansion.
• English tool names with Arabic descriptions — models trained primarily on English function-calling may have a spurious advantage on function selection. See docs/grading.md.
• No automatic audio, RTL-UI, or code-switching tests. Text-only evaluation.
• Grading normalizes Arabic variants conservatively (alef, ya, ta-marbuta, tatweel) but does not apply deep lemmatization. Edge cases may score lower than a human would award.