ArSyra/arsyra-nlp-benchmark

📊 ArSyra NLP Benchmark — Arabic Dialect Evaluation Suite

The first Arabic NLP benchmark that spans dialects, not just MSA.

---

---

Dataset Summary

A structured evaluation dataset for benchmarking Arabic NLP models on dialect-aware tasks. Contains sentiment-annotated text, quality control labels with human judgments, and instruction-description pairs for testing model comprehension and generation capabilities.

Unlike most Arabic benchmarks that focus exclusively on MSA, ArSyra NLP Benchmark spans multiple dialect groups, enabling fair evaluation of how well models handle the Arabic that people actually speak.

Statistic	Value
Total Records	2,419
Linguistic Categories	3
Countries Represented	15 (Tunisia, Syria, Egypt, Saudi Arabia, Morocco, Algeria, Iraq, Jordan, Lebanon, UAE, Sudan, Yemen, Libya, Kuwait, Palestine)
Dialect Groups	7 (Maghrebi, Levantine, Egyptian, Gulf, Iraqi, Sudanese, Other)
Average Quality Score	78.2/100
License	CC-BY-NC-SA-4.0
Last Updated	2026-02-21

How ArSyra Compares to Existing Arabic Datasets

Dataset	Records	Dialects	Countries	Categories	Crowdsourced	MSA↔Dialect Pairs
ArSyra (arsyra-nlp-benchmark)	2,419	7	15	3	✅	✅
NADI (shared task)	~20K	4	21	1	❌ (Twitter)	❌
MADAR	~12K	6	25	1	✅ (paid)	✅
AOC (Arabic Online Commentary)	~100K	—	—	3	❌ (scraped)	❌
DART (Dialect Arabic)	~25K	5	—	1	❌ (Twitter)	❌
ArSentD-LEV	~4K	1	4	1	❌ (Twitter)	❌

ArSyra's advantages: Authentic native-speaker data (not scraped), multi-category structure, parallel MSA↔dialect text, quality scored, and continuously growing.

Related ArSyra Datasets

Explore our other specialized Arabic dialect datasets:

• 🏆 ArSyra Complete — Multi-Dialect Arabic Dataset — The most comprehensive crowdsourced Arabic dialect dataset available.
• 🤖 ArSyra Chatbot — Conversational Arabic Training Data — Purpose-built training data for Arabic conversational AI systems.
• 🌐 ArSyra Translation — Arabic Dialect–MSA Parallel Corpus — Parallel corpus bridging Modern Standard Arabic and regional dialects.
• 🇪🇬 ArSyra Egyptian Arabic (Masri) Dataset — The most widely understood Arabic dialect — now as structured NLP data.
• 🇸🇾 ArSyra Levantine Arabic (Shami) Dataset — Authentic Shami dialect data from Syria, Lebanon, Jordan, and Palestine.
• 🇸🇦 ArSyra Gulf Arabic (Khaliji) Dataset — Gulf Arabic data from the Arabian Peninsula's rapidly growing digital population.

Browse all datasets: huggingface.co/ArSyra | arsyra.com/datasets.html

Supported Tasks

• Text Classification — Train classifiers for dialect identification, sentiment analysis, and content categorization.
• Text Generation — Fine-tune language models to generate authentic dialectal Arabic text.

Languages

Primary Language: Arabic (ar)

This dataset contains text in Modern Standard Arabic (MSA) and the following regional dialect groups: Maghrebi, Levantine, Egyptian, Gulf, Iraqi, Sudanese, Other. Country-level dialect codes: ar-TN, ar-SY, ar-EG, ar-SA, ar-MA, ar-DZ, ar-IQ, ar-JO, ar-LB, ar-AE, ar-SD, ar-YE, ar-LY, ar-KW, ar-PS.

---

Dataset Structure

Data Instances

Each record represents a single response from a verified native Arabic speaker to a structured linguistic prompt:

{
  "question_code": "I-0015",
  "category": "instructions",
  "subcategory": "food",
  "question_text": "اشرح كيف تذبح خروف أو دجاجة بلهجتك (خطوة بخطوة)",
  "answer_text": "تقبل القبلة وتسمي بسم الله وتذبح",
  "response_time_ms": 84609,
  "quality_score": 100,
  "country": "TN",
  "answered_at": "2026-02-17T21:15:07.495Z",
  "quality_grade": "A",
  "speaker_hash": "anon-d2ViLTE3"
}

Data Fields

Field	Type	Description
text	string	The Arabic text content — may be in dialect, MSA, or a mix
category	string	Linguistic category (e.g., dialect, proverbs, sentiment, conversation_pairs)
country	string	ISO 3166-1 alpha-2 country code of the speaker (e.g., EG, SA, MA)
dialect_group	string	Broad dialect group: egyptian, levantine, gulf, maghrebi, iraqi, or sudanese
quality_score	int	Human-assigned quality rating from 0 to 100
msa_text	string	Modern Standard Arabic equivalent (where available)
context	string	Additional context about the prompt or response
speaker_hash	string	Anonymized speaker identifier

Data Splits

Split	Examples
train	2,419

Note: A single train split is provided. We recommend creating your own train/validation/test splits based on your use case. For dialect-fair evaluation, stratify by country or dialect_group.

Category Breakdown

Category	Records	% of Total
instructions	1,400	57.9%
sentiment	719	29.7%
control	300	12.4%

---

Dataset Creation

Curation Rationale

Existing Arabic NLP benchmarks (ORCA, ALUE) focus almost exclusively on MSA text, creating a misleading picture of model capabilities. Arabic speakers primarily communicate in dialect, and models need to be evaluated accordingly. ArSyra NLP Benchmark fills this gap.

Source Data

Initial Data Collection and Normalization

Data was collected through the ArSyra platform (arsyra.com), a gamified crowdsourcing system where verified native Arabic speakers answer structured linguistic prompts about their dialect. The platform:

1. Verifies speakers through phone number verification (region-specific) and language verification questions
2. Presents structured prompts across multiple linguistic categories: dialect translations, conversation pairs, proverbs, slang, code-switching, sentiment expressions, instruction following, formality registers, and more
3. Gamifies collection through points, leaderboards, and incentive systems to maintain engagement and data quality
4. Automatically enriches responses with metadata: country, dialect group, category, and quality indicators

Who are the source language producers?

Native Arabic speakers from 15 countries across the Arab world (Tunisia, Syria, Egypt, Saudi Arabia, Morocco, Algeria, Iraq, Jordan, Lebanon, UAE, Sudan, Yemen, Libya, Kuwait, Palestine), participating voluntarily through the ArSyra platform. Speakers represent diverse demographics including age groups, education levels, and urban/rural backgrounds.

Annotations

Annotation Process

Each response receives:

• Automatic quality scoring based on response length, character set validation, and consistency checks
• Category labeling derived from the prompt type
• Dialect group classification based on the speaker's registered country
• Cross-speaker validation where multiple speakers from the same region answer the same prompts

Who are the annotators?

The primary "annotators" are the native speakers themselves, who provide dialectal data along with structured metadata. Quality scoring is automated. No external annotators are used for labeling.

Personal and Sensitive Information

• All speaker identifiers are anonymized — original user IDs are replaced with non-reversible hashed identifiers
• No personally identifiable information (names, locations, phone numbers) is included
• Taboo and sensitive content (where present) is clearly labeled by category
• Speakers provided informed consent during registration for their anonymized data to be used for research

---

Considerations for Using the Data

Social Impact

This dataset contributes to Arabic NLP equity by providing training data for the dialects actually spoken by 400+ million people. Most existing Arabic NLP resources focus exclusively on Modern Standard Arabic, which is no one's native language. By bridging this gap, ArSyra helps ensure that Arabic-speaking populations benefit equally from advances in language technology.

Discussion of Biases

Known biases to consider:

1. Platform access bias — Contributors need internet access and a smartphone, potentially underrepresenting older, rural, or lower-income speakers
2. Country representation — Some countries may be overrepresented depending on recruitment channels
3. Urban bias — Online populations tend to be more urban, potentially underrepresenting rural dialect variants
4. Literacy bias — Written responses may differ from purely spoken dialect, as speakers may unconsciously shift toward MSA
5. Self-selection bias — Voluntary participants may not represent the full demographic spectrum

Other Known Limitations

• Written approximations — Dialectal Arabic has limited standardized orthography; spelling varies across speakers
• Prompt influence — Structured prompts may elicit more formal responses than spontaneous speech
• Quality variation — Despite quality scoring, some responses may be lower quality
• Temporal snapshot — Language evolves; slang and expressions may become dated over time

---

Additional Information

Use Cases

• Benchmarking Arabic LLMs on dialectal understanding
• Evaluating sentiment analysis across dialect groups
• Testing instruction-following in non-MSA Arabic
• Comparing model performance across Arabic varieties

Get the Full Dataset

This repository contains a preview sample of 50 records out of 2,419 total. Purchase the full dataset instantly at arsyra.com/datasets.html

Pricing

Preview (this repo)	50 sample records — free to download and evaluate
Full Dataset	2,419 records — instant download after purchase
Academic License	From $29 — for research and non-commercial use
Commercial License	From $99 — for products, SaaS, and enterprise use

🛒 Buy Now →

What you get with the full dataset:

• All 2,419 quality-filtered records
• Per-category JSONL splits for easy loading
• Instant download as ZIP after payment
• Regular updates as our community grows
• Priority support for integration questions

Questions? Email support@arsyra.com

---

Quick Start

from datasets import load_dataset

# Load the preview sample
dataset = load_dataset("ArSyra/arsyra-nlp-benchmark")
print(f"Preview: {len(dataset['train'])} sample records")

# Browse examples
for example in dataset["train"].select(range(5)):
    print(f"{example['country']} ({example['dialect_group']}): {example['text'][:80]}...")

# For the full dataset (2,419 records), visit: https://arsyra.com/datasets.html

Licensing Information

The preview sample included in this repository is released under CC-BY-NC-SA-4.0.

The full dataset is available under flexible licensing terms:

License	Use Case	Pricing
CC-BY-NC-SA-4.0	Academic research, non-commercial use	From $29
Commercial License	Enterprise, products, SaaS applications	From $99

Purchase a license → or email support@arsyra.com for custom licensing.

Citation Information

If you use this dataset in your research, please cite:

@dataset{arsyra_arsyra_nlp_benchmark_2026,
  title     = {ArSyra NLP Benchmark — Arabic Dialect Evaluation Suite},
  author    = {{ArSyra Team}},
  year      = {2026},
  url       = {https://huggingface.co/datasets/ArSyra/arsyra-nlp-benchmark},
  publisher = {HuggingFace},
  license   = {CC-BY-NC-SA-4.0},
  note      = {Crowdsourced Arabic dialect dataset with 2,419 records from 15 countries}
}

Contributions

Thanks to the Arabic-speaking community who contributed their dialectal knowledge through the ArSyra platform. To contribute, visit arsyra.com.

---

Dataset card generated by the ArSyra Publish Pipeline. Last updated: 2026-02-21.