Datasets:
Download README.md from alaabaz/MuDAS: direct link, hf CLI and curl.
- Browser
- Download file 8.17 kB
-
https://huggingface.co/datasets/alaabaz/MuDAS/resolve/main/README.md
- Command line
-
hf download hf://datasets/alaabaz/MuDAS/README.md
-
curl -L -H "Authorization: Bearer $HF_TOKEN" -o README.md https://huggingface.co/datasets/alaabaz/MuDAS/resolve/main/README.md
language:
- ar
license: cc-by-nc-4.0
pretty_name: 'MuDAS: Multi-Dialect Arabic Sexism Dataset'
task_categories:
- text-classification
tags:
- sexism-detection
- arabic
- dialects
- annotator-disagreement
- soft-labels
- perspectivism
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files: MuDAS_dataset.csv
extra_gated_heading: Request access to MuDAS
extra_gated_prompt: >-
MuDAS contains YouTube comments that may be offensive, sexist or harmful.
Access is granted for non-commercial research only. Requests are reviewed
manually.
extra_gated_button_content: Submit request
extra_gated_fields:
Full name: text
Affiliation: text
Academic / institutional email: text
Intended use of the dataset: text
I will use this dataset for non-commercial research purposes only: checkbox
I will not redistribute the dataset or any part of it: checkbox
I will not attempt to identify the authors of the comments or the annotators: checkbox
I will cite the MuDAS paper in any work that uses this dataset: checkbox
MuDAS: Multi-Dialect Arabic Sexism Dataset
⚠️ Content warning: this dataset contains comments that are sexist, offensive or otherwise harmful.
Dataset summary
MuDAS is a dataset of 6,731 Arabic YouTube comments annotated for sexism in three Arabic dialects: Saudi, Egyptian and Moroccan. Each comment was labelled independently by three native speakers of the comment's dialect. The dataset releases every annotator's label, not only the majority label, so it supports hard-label, soft-label (disagreement-aware) and perspectivist modelling, as well as the analysis of how judgements of sexism vary within and across dialect communities.
MuDAS is introduced in the paper What Majority Labels Hide: Annotator Disagreement and Model Errors in Multi-Dialect Arabic Sexism Detection (see Citation).
Dataset statistics
| Dialect | Comments | Sexist (majority) | Not sexist (majority) | % sexist |
|---|---|---|---|---|
| Saudi | 2,257 | 963 | 1,294 | 42.7% |
| Egyptian | 2,149 | 742 | 1,407 | 34.5% |
| Moroccan | 2,325 | 1,145 | 1,180 | 49.2% |
| Total | 6,731 | 2,850 | 3,881 | 42.3% |
Number of annotators (out of three) who labelled a comment as sexist:
| Dialect | 0 of 3 | 1 of 3 | 2 of 3 | 3 of 3 |
|---|---|---|---|---|
| Saudi | 854 | 440 | 413 | 550 |
| Egyptian | 1,083 | 324 | 311 | 431 |
| Moroccan | 863 | 317 | 493 | 652 |
Comments labelled sexist by one or two of the three annotators are contested; comments labelled sexist by none or all three are unanimous.
Data splits
The paper used the following split. The test sets were drawn by stratified sampling on dialect and agreement level, and 15% of the remaining comments were held out for validation.
| Dialect | Train | Validation | Test | Total |
|---|---|---|---|---|
| Saudi | 1,534 | 271 | 452 | 2,257 |
| Egyptian | 1,461 | 258 | 430 | 2,149 |
| Moroccan | 1,581 | 279 | 465 | 2,325 |
| Total | 4,576 | 808 | 1,347 | 6,731 |
Data fields
| Column | Description |
|---|---|
UID |
Unique comment ID: <dialect>_W<batch>_S<session>_Q<item> (e.g. saudi_W3_S2_Q15) |
Dialect_Group |
Dialect of the comment and its annotators: Saudi, Egyptian, Moroccan |
Video_Description |
Short neutral description (in Arabic) of the YouTube video the comment was posted under, shown to annotators as context |
Comment |
Comment text |
Sexism_A1, Sexism_A2, Sexism_A3 |
Each annotator's label (Yes = sexist, No = not sexist) |
Sexism_Majority |
Majority label across the three annotators |
Sexism_Soft |
Share of annotators who labelled the comment sexist (0, 0.33, 0.67, 1) |
Sexism_Agreement |
Share of annotators agreeing with the majority label (0.67 or 1) |
Dialect_Majority |
Whether the majority of annotators confirmed the comment is written in the target dialect |
Dialect_Agreement |
Share of annotators agreeing with the dialect majority |
Important: A1, A2 and A3 are different people in each dialect. For example, Sexism_A1 for Saudi comments and Sexism_A1 for Egyptian comments were produced by different annotators. Within one dialect, each column always refers to the same annotator.
Data collection
- Source: public comments (top-level comments and replies) from YouTube videos on gender-related topics, collected separately for each dialect region. Videos were found with 32 Arabic search keywords grouped into five themes (marriage and divorce, work and independence, appearance and behaviour, roles and relationships, controversial issues).
- Videos: the list of selected videos, with their theme and dialect, is provided in
MuDAS_videos.csv. - Pre-screening: comments were pre-screened with Llama-3.1-8B-Instruct (Arabic prompt), and comments predicted as sexist were oversampled. The dataset is therefore enriched for likely sexist content and does not reflect the natural prevalence of sexism on YouTube.
- Anonymisation: user names and channel information were removed; user mentions were replaced with
@USER, and URLs and email addresses withURLandEMAIL.
Annotation
- Annotators were recruited through Prolific and screened for nationality, birthplace and residence matching the dialect.
- Each dialect was labelled by the same three annotators (two men and one woman in each dialect), all of whom passed a qualification pilot.
- Annotators followed dialect-specific versions of the annotation guidelines (Saudi, Egyptian and Moroccan), with a definition of sexism and examples of sexist and non-sexist comments.
- The task used a binary label (sexist / not sexist). Annotators also confirmed whether each comment was written in their dialect; comments not confirmed as such were removed.
- Quality control: each session included hidden gold items, real comments whose label follows clearly from the guidelines, and from the fourth batch one comment from an earlier batch was repeated in each session to measure self-consistency. Individual accuracy on the gold items ranged from 95.3% to 100%, and annotators agreed with their earlier label on 93.4% of the repeated comments on average. Repeated comments were reduced to their first occurrence.
- Annotators were paid £15 per hour.
- Inter-annotator agreement (Fleiss' κ) is reported in the paper.
Intended uses
- Arabic sexism detection (hard labels)
- Learning from annotator disagreement (soft labels)
- Perspectivist modelling (individual annotator labels)
- Analysis of how judgements of sexism vary across annotators and dialects
Out-of-scope uses: commercial use, identifying or profiling the authors of the comments or the annotators, or generating sexist content.
Limitations
- Each dialect was labelled by the same three annotators, who cannot represent the range of views within their community; differences between dialects cannot be separated from differences between annotator groups.
- All comments come from YouTube videos on gender-related topics, and the data are enriched for likely sexist content.
- The label is binary and does not capture which aspects of a comment the annotators disagreed on.
Ethics
This research was approved by the Queen Mary University of London research ethics committee (ref. DSEECS26.020). All annotators gave informed consent, were informed that the task involved offensive content, and could withdraw at any time.
Takedown: if you are the author of a comment in this dataset and would like it removed, please contact us and we will remove it.
Citation
@misc{bazaid2026majoritylabelshideannotator,
title={What Majority Labels Hide: Annotator Disagreement and Model Errors in Multi-Dialect Arabic Sexism Detection},
author={Alaa Bazaid and Arkaitz Zubiaga},
year={2026},
eprint={2611.24383},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2611.24383},
}
Contact
Alaa Bazaid, Queen Mary University of London: a.z.o.bazaid@qmul.ac.uk