---
license: cc-by-4.0
task_categories:
  - translation
language:
  - ar
  - en
tags:
  - sudanese-arabic
  - dialectal-arabic
  - machine-translation
  - benchmark
  - low-resource-nlp
size_categories:
  - n<1K
---

# Sudanese Arabic Machine Translation Benchmark

## Dataset Description

This benchmark addresses a **documented gap** in Arabic NLP research: the exclusion of Sudanese Arabic from dialectal machine translation evaluation.

Recent research (Alabdullah et al., 2025) states:
> "The focus on three dialects (Levantine, Gulf, Egyptian) constrained generalization, leaving the proposed translation techniques **untested on varieties such as Maghrebi and Sudanese Arabic**."

This dataset provides the first benchmark specifically designed for evaluating Sudanese Arabic → English translation quality.

## Supported Tasks

- **Translation**: Sudanese Arabic → English
- **Dialect Identification**: Can be used to test if models distinguish Sudanese from other Arabic dialects
- **Error Analysis**: Annotated for linguistic challenges unique to Sudanese

## Languages

- **Source**: Arabic (Sudanese dialect)
- **Target**: English

## Dataset Structure

### Data Fields

| Field | Type | Description |
|-------|------|-------------|
| id | string | Unique identifier (sud_mt_XX) |
| source | string | Sudanese Arabic text |
| reference | string | English reference translation |
| difficulty | string | easy/medium/hard |
| notes | string | Linguistic annotation |
| dialect_markers | list | Sudanese-specific features present |
| confusion_risk | string | Potential for dialect confusion |

### Data Splits

- **test**: 20 sentences (10 Sudanese Arabic + 10 MSA control)

## Linguistic Features

### Unique Sudanese Features

| Feature | Arabic | Meaning | MT Impact |
|---------|--------|---------|-----------|
| يا زول | ya zol | "hey man" | High - unique vocative |
| حي- | hay- | future prefix | High - unique morphology |
| يتعرس | yit'arris | gets married | High - unique verb |
| حتة | hita | room | Medium - context-dependent |

### Features Shared with Egyptian (Confusion Risk)

| Feature | Shared Meaning | Confusion Potential |
|---------|----------------|---------------------|
| دي | "this" | High |
| عايز | "want" | High |
| مش | "not" | Medium |

## Citation

```bibtex
@dataset{mihaysi2026sudanese_mt_benchmark,
  title = {Sudanese Arabic Machine Translation Benchmark},
  author = {Mihaysi, Aamer},
  year = {2026},
  publisher = {HuggingFace},
  note = {Addresses gap identified in Alabdullah et al. (2025)}
}
```

## References

- Alabdullah, A., Han, L., & Lin, C. (2025). Advancing Dialectal Arabic to Modern Standard Arabic Machine Translation. arXiv:2507.20301.
- Beidas, A., et al. (2025). Cross-dialectal Arabic translation: comparative analysis on large language models. Frontiers in Artificial Intelligence.

## Related Datasets

- [O96a/sudanese-arabic-dialect-benchmark](https://huggingface.co/datasets/O96a/sudanese-arabic-dialect-benchmark) - Dialect identification benchmark
- [O96a/opus-mt-arabic-benchmark-2026-03-28](https://huggingface.co/datasets/O96a/opus-mt-arabic-benchmark-2026-03-28) - Arabic MT benchmark

## License

Creative Commons Attribution 4.0 (CC-BY-4.0)