O96a/sudanese-arabic-dialect-benchmark

Sudanese Arabic Dialect Identification Benchmark

Dataset Summary

A curated benchmark dataset for Arabic dialect identification, with special focus on Sudanese Arabic (ARS) — an under-represented variety in NLP research.

This dataset enables evaluation of:

• Dialect identification models
• Cross-dialectal transfer learning
• Low-resource language modeling
• Arabic NLP robustness testing

Supported Tasks and Leaderboards

Task	Description	Metric
text-classification	Identify dialect from text	Accuracy, F1
zero-shot-classification	Classify without training	Accuracy
translation-quality	Evaluate MT on dialects	BLEU, chrF

Languages

• Primary: Sudanese Arabic (ARS)
• Comparison: MSA (Modern Standard Arabic), Egyptian Arabic (ARZ), Levantine Arabic (APC), Gulf Arabic (ABG)

---

Dataset Structure

Data Instances

{
  "id": "sud_001",
  "text": "يا زول، كيف حالك؟ تعال نتغدا سوا اليوم",
  "dialect": "sudanese",
  "length": "short"
}

Data Fields

Field	Type	Description
id	string	Unique identifier (prefix indicates dialect)
text	string	Arabic text sample
dialect	string	Dialect label: sudanese, msa, egyptian, levantine, gulf
length	string	Text length category: short (≤10 words), medium (11-20 words)

Data Splits

Split	Samples	Sudanese	MSA	Egyptian	Levantine	Gulf
train	80	20	16	16	16	16
test	20	5	4	4	4	4
Total	100	25	20	20	20	20

---

Dataset Creation

Curation Rationale

Sudanese Arabic is significantly under-represented in Arabic NLP research. Most dialect identification benchmarks focus on Egyptian, Levantine, and Gulf varieties, leaving Sudanese Arabic unaddressed.

This dataset fills that gap by providing:

1. Minimal pairs — Semantically equivalent sentences across dialects
2. Real corpus samples — Authentic Sudanese Arabic from the private O96a/sudanese-dialect corpus (464K+ samples)
3. Balanced distribution — Equal representation across dialects

Source Data

Initial Data Collection

Dialect	Source	Method
Sudanese	O96a/sudanese-dialect corpus	Random sampling + curation
MSA	Curated	Native speaker generation
Egyptian	Curated	Native speaker generation
Levantine	Curated	Native speaker generation
Gulf	Curated	Native speaker generation

Who are the source language producers?

• Sudanese: Social media posts, personal communications (Sudanese Arabic speakers)
• Other dialects: Native speakers of each dialect

Annotations

Annotation process

1. Text samples collected/curated
2. Dialect labels assigned by native speakers
3. Length categorization (short/medium)
4. Quality review for naturalness

Annotation guidelines

• Use authentic, natural expressions
• Avoid code-switching (pure Arabic)
• Include common conversational patterns
• Ensure semantic equivalence across dialects

Who are the annotators?

Aamer Mihaysi (O96a) — native Sudanese Arabic speaker, NLP researcher

---

Considerations in Using the Data

Social Impact of Dataset

This dataset enables:

• Better representation of African Arabic varieties in NLP
• Development of dialect-aware Arabic language technologies
• Preservation and documentation of Sudanese Arabic

Limitations

1. Size — Limited to 100 samples (suitable for evaluation, not training)
2. Domain — Conversational/social media text; may not generalize to formal domains
3. Coverage — 5 dialects only; excludes Maghrebi, Iraqi, Yemeni, etc.
4. Annotation — Single annotator; no inter-annotator agreement

Bias and Risks

• Samples may reflect particular demographics (young, social media users)
• Sudanese samples may include regional sub-varieties (Khartoum, Omdurman, etc.)
• Some dialect markers may overlap (e.g., "وين رايح" appears in both Sudanese and Levantine)

Ethical Considerations

• All Sudanese corpus samples are anonymized
• No personally identifiable information
• Intended for research purposes only

---

Additional Information

Dataset Curators

Aamer Mihaysi (O96a)

• HuggingFace: https://huggingface.co/O96a
• Website: https://www.mehaisi.com
• Organization: Sudaverse

Licensing Information

Apache License 2.0

Free for academic and commercial use with attribution.

Citation Information

@dataset{mihaysi2026sudanese_benchmark,
  title     = {Sudanese Arabic Dialect Identification Benchmark},
  author    = {Mihaysi, Aamer},
  publisher = {HuggingFace},
  year      = {2026},
  url       = {https://huggingface.co/datasets/O96a/sudanese-arabic-dialect-benchmark}
}

Contributions

Thanks to the Sudanese Arabic NLP community for inspiration and feedback.

Related Resources

• Private Corpus: O96a/sudanese-dialect — 464K+ Sudanese Arabic text samples
• Previous Benchmark: O96a/opus-mt-arabic-benchmark-2026-03-28 — Translation quality evaluation

---

Sudanese Arabic Linguistic Features

Key Dialect Markers

Feature	Sudanese	MSA	Egyptian	Levantine
"Friend/Guy"	يا زول (ya zol)	يا صديقي	يا صاحبي	يا خي
"What"	شنو (sheno)	ماذا	إيه	شو
"Today"	اليوم (al-yom)	اليوم	النهاردة	اليوم
"Room"	حتة (huta)	غرفة	أوضة	غرفة
"This"	ده (de)	هذا	ده	هادا/هالـ

Morphological Features

• Negation: ما + verb (ما قدرت → "couldn't")
• Question particles: شنو, وين
• Demonstratives: ده, دي (this, this-fem)

Syntax Notes

• Subject-Verb-Object (SVO) common in spoken varieties
• Reduced case marking compared to MSA
• Frequent use of discourse particles (والله, يا زول)