Sudan-MM: A Multimodal Dataset of Sudanese Arabic

Sudan-MM is the first publicly available multimodal dataset for Sudanese Arabic (السودانية), a low-resource dialect with no prior paired image-caption, video-caption, or voice-caption data. It was produced through a competitive shared task held in 2025, where five teams collected and annotated media depicting everyday Sudanese life.

Each item in the dataset pairs a visual or video recording with:

• a written caption in Modern Standard Arabic (MSA)
• a written caption in Sudanese Arabic dialect
• a spoken audio recording of the Sudanese Arabic caption

---

Dataset at a Glance

	Images	Videos	Total
Items	2,742	300	3,042
With audio	1,842	187	2,029
Contributing teams	5	5	5
Categories	9	9	9

Audio coverage varies by team; SlothYouth submitted audio for images only (101 of 1,000).

---

Data Fields

images.csv

Field	Type	Description
image_id	string	Canonical ID, e.g. Image_0001
image_path	string	Relative path to the image file
team	string	Contributing team name
audio_path	string	Relative path to the MP3 voice caption (may be empty)
caption_MSA	string	Caption in Modern Standard Arabic
caption_sudanese	string	Caption in Sudanese Arabic dialect
category	string	One of 9 canonical thematic categories (see below)
secondary_category	string	Optional finer-grained sub-label supplied by the team

videos.csv

Same schema with video_id / video_path instead of image_id / image_path.

---

Categories

After normalization across all team submissions, items are assigned to one of 9 canonical categories:

Category	Items
Urban & City Life	381
Nature & Landscape	341
Public Spaces & Infrastructure	267
Local Culture & Objects	225
Agriculture & Livestock	221
Food & Drinks	179
Rural & Daily Life	115
Marketplaces	100
Transportation	100
Uncategorized (SlothYouth — no labels submitted)	1,113

---

Per-Team Statistics

Team	Images	Videos	Audio (img)	Audio (vid)	Final Score /100
MalamhSudan (1st)	402	42	402	42	84.29
4sparks (2nd)	640	55	639	55	80.58
معالم في الطريق (3rd)	300	100	300	50	74.73
Hope (4th)	400	40	400	40	73.82
SlothYouth (5th)	1,000	63	101	0	46.57

---

Evaluation

Submissions were scored out of 100 points across three components:

Automated (50 pts) — heuristic code checks: folder structure, ID/naming compliance, metadata completeness, file validity (format, duration, corruption, blur), and category diversity metrics.

LLM quality scoring (40 pts) — a stratified random sample of 30 items per team was scored by Claude Haiku on MSA caption quality, Sudanese dialect authenticity, and category correctness. Media and audio quality were derived from automated findings.

Human paper review (10 pts) — two organiser judges independently rated each team's overall submission on five rubric criteria (data quality, diversity, annotation quality, documentation, originality) plus a holistic overall score. Scores were z-score normalised per reviewer and combined 70/30 criteria/overall.

---

Dataset Creation

Task Design

Teams were required to submit per-item:

• An original image (JPG/PNG) or short video (MP4, 3–10 s)
• An MSA Arabic written caption
• A Sudanese Arabic dialect written caption
• A spoken MP3 audio recording of the Sudanese caption (5–15 s)
• A thematic category label from a predefined list

Detailed technical requirements are available in the Technical Requirements Document.

Source Data

All media was collected and annotated by the five competing teams. Content depicts everyday Sudanese life: food, markets, transportation, landscapes, cultural objects, agriculture, and urban/rural scenes.

Known Limitations

• SlothYouth submitted no category labels and minimal audio, resulting in 1,113 uncategorised items and lower quality scores.
• Category labels were submitted independently by each team using inconsistent terminology; all have been normalised to the 9-class taxonomy above.
• Audio coverage is not 100% across all teams.
• Some MSA captions are missing for a small number of video items (particularly in the 4sparks submission).

---

Languages

Code	Name
ar	Modern Standard Arabic (MSA)
ar-SD	Sudanese Arabic dialect

Sudanese Arabic differs substantially from MSA and other Arabic dialects in vocabulary, phonology, and morphology. It is spoken by approximately 45 million people and has no standardised written form.

---

Citation

If you use Sudan-MM in your research, please cite:

@dataset{sudanmm2025,
  title     = {Sudan-MM: A Multimodal Dataset of Sudanese Arabic},
  author    = {Sudan-MM Organizers},
  year      = {2025},
  license   = {Apache-2.0},
}

---

License

This dataset is released under the Apache 2.0 License.