Hugging Face's logo

Datasets:
TeamSmart
/
BOUTEF

Modalities:
Tabular
Text
Formats:
csv
Languages:
Arabic
French
English
Size:
100K - 1M
ArXiv:
Tags:
Fake
Disinformation
Arabic
Dialect
Algerian
Tunisian
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

By requesting access to BOUTEF, I confirm that:

  • I will use the dataset only for non-commercial research or educational purposes.
  • I will not redistribute, mirror, republish, sublicense, sell, transfer, or publicly expose the dataset.
  • I will not attempt to identify, re-identify, contact, track, profile, or deanonymize users, accounts, authors, or commenters.
  • I will not combine the dataset with external resources in a way that could enable re-identification.
  • I will not use the dataset for advertising, marketing, user profiling, political targeting, surveillance, law-enforcement profiling, harassment, discrimination, or misinformation generation.
  • I will not publish raw examples, screenshots, links, metadata, or model outputs that could reveal user or account identities.
  • I will cite the associated BOUTEF dataset paper in any publication or derived work.
  • I accept the Controlled Research Use Data Agreement.

Log in or Sign Up to review the conditions and access this dataset content.

BOUTEF: A Multilingual Corpus for Fake News in North Africa

BOUTEF is a controlled-access research dataset for the study of misinformation in North African social media. It contains misinformation narratives, genuine reference items, user-generated reactions, and debunking or verification-related material collected from online sources and social media platforms.

The dataset includes multilingual and multidialectal content, including Modern Standard Arabic, Algerian dialect, Tunisian dialect, Moroccan dialect, Arabizi, French, English, and code-switched text.

Access to this dataset is restricted. By requesting, accessing, downloading, or using the dataset, users agree to the Controlled Research Use Data Agreement described below.

Dataset Access

This dataset is distributed under a Controlled Research Use Data Agreement.

Access is granted only for non-commercial research and educational purposes. Users must not redistribute, mirror, republish, sublicense, sell, transfer, or expose the dataset through public APIs, public demos, or public applications.

Researchers requesting access should confirm that they have read and accepted the usage checklist below.

Permitted Uses

The dataset may be used only for non-commercial research and educational purposes.

Permitted uses include:

  • academic research;
  • benchmarking and evaluation;
  • reproducibility studies;
  • development of methods for misinformation detection;
  • multilingual and dialectal NLP research;
  • analysis of user reactions, debunking, and misinformation discourse;
  • responsible evaluation of NLP systems on North African social-media content.

Prohibited Uses

Researchers shall not use the dataset for:

  • commercial product development or paid services;
  • advertising, marketing, or user profiling;
  • political targeting, persuasion, or influence operations;
  • surveillance or monitoring of individuals or communities;
  • law-enforcement profiling;
  • harassment, discrimination, or targeting of individuals or groups;
  • generation, amplification, or optimization of misinformation;
  • attempts to infer sensitive personal attributes of users;
  • any use that may cause harm to individuals, communities, or vulnerable groups.

Controlled Access and Redistribution

Access to the dataset is granted on an individual basis.

Researchers shall not:

  • redistribute the dataset;
  • mirror or republish the dataset;
  • upload the dataset to public repositories;
  • sublicense, sell, or transfer the dataset;
  • share the dataset with unauthorized users;
  • expose the dataset through public APIs, demos, or applications.

The dataset may not be uploaded to third-party services unless those services provide equivalent access control, confidentiality, and data-protection safeguards.

The dataset providers reserve the right to approve, deny, suspend, or terminate access at any time.

Research Team Access

Researchers may provide access to students, research assistants, or collaborators working under their direct supervision only if each person has agreed to be bound by these terms.

The primary researcher remains responsible for ensuring that all collaborators comply with this agreement.

Access must not be shared with individuals outside the approved research team.

Privacy and Non-Identification

The dataset may contain user-generated content, social-media reactions, and platform-derived metadata.

Researchers shall not attempt to:

  • identify;
  • re-identify;
  • contact;
  • track;
  • profile;
  • deanonymize;

any individual, author, commenter, account, or social-media user represented in or associated with the dataset.

Researchers shall not combine the dataset with external resources in a way that could reasonably enable re-identification.

Researchers shall not publish raw examples, screenshots, links, metadata, or model outputs that could reveal the identity of users, accounts, or private individuals.

Anonymisation and Data Cleaning

The distributed version of BOUTEF is an anonymized version of the collected dataset.

The anonymization process consists in replacing the username in USER field by a random string. When the username is associated in the USER field with the full name, then, this full name is also replaced by the same randomized string. For example, if the USER is "John Smith @smith1234", then, smith1234 could be replaced by ajrbodft and then John Smith is replaced by appelation_ajrbodft. This replacement process is done on the whole corpus, including USER and MESSAGE fields: in the whole corpus, smith1234 is replaced by ajrbodft and John Smith is replaced by appelation_ajrbodft.

Dataset Structure

The dataset is provided as an XML file containing a root <Entries> element. Each data item is represented by an <Entry> element.

A simplified structure is shown below:

<Entries>
  <Entry>
    <CONCERN>ALGERIE</CONCERN>
    <TYPE>CHECKED</TYPE>
    <MAJOR_THEME>LANGUE</MAJOR_THEME>
    <MINOR_THEME></MINOR_THEME>
    <SUBJECT>1991 MANIF ALGERIE FRANCAIS</SUBJECT>
    <IMAGE_REFERENCE></IMAGE_REFERENCE>
    <MESSAGE>...</MESSAGE>
    <TYPE_MESSAGE>FAKE</TYPE_MESSAGE>
    <FAKE_CATEGORY>MISLEADING CONTENT</FAKE_CATEGORY>
    <LANGUAGE>MSA</LANGUAGE>
    <PUBLICATION_DATE>2023-04-02 00:00:00</PUBLICATION_DATE>
    <COMMENTS_NUMBER></COMMENTS_NUMBER>
    <SITE>Twitter</SITE>
    <NUMBER_FOLLOWERS>17300</NUMBER_FOLLOWERS>
    <NUMBER_OF_FORWARD></NUMBER_OF_FORWARD>
    <COUNTRY></COUNTRY>
    <GENDER>UND</GENDER>
    <USER>appelation_xxxxxxxx @xxxxxxxx</USER>
  </Entry>

  <Entry>
    <CONCERN>ALGERIE</CONCERN>
    <TYPE>CHECKED</TYPE>
    <MAJOR_THEME>LANGUE</MAJOR_THEME>
    <MINOR_THEME></MINOR_THEME>
    <SUBJECT>1991 MANIF ALGERIE FRANCAIS</SUBJECT>
    <IMAGE_REFERENCE></IMAGE_REFERENCE>
    <MESSAGE>...</MESSAGE>
    <TYPE_MESSAGE>FAKECOMMENT</TYPE_MESSAGE>
    <FAKE_CATEGORY></FAKE_CATEGORY>
    <LANGUAGE>MSA</LANGUAGE>
    <PUBLICATION_DATE>2023-04-07 00:00:00</PUBLICATION_DATE>
    <COMMENTS_NUMBER></COMMENTS_NUMBER>
    <SITE>Twitter</SITE>
    <NUMBER_FOLLOWERS></NUMBER_FOLLOWERS>
    <NUMBER_OF_FORWARD></NUMBER_OF_FORWARD>
    <COUNTRY></COUNTRY>
    <GENDER>UND</GENDER>
    <USER>appelation_yyyyyyyy @yyyyyyyy</USER>
  </Entry>
</Entries>

The XML example above is illustrative. Researchers must not publish raw dataset entries, screenshots, user identifiers, URLs, or platform metadata that could enable re-identification.

Field Description

Field Description
CONCERN Country, region, or geopolitical context concerned by the item.
TYPE Verification status or source type, for example checked content.
MAJOR_THEME Broad thematic category of the item.
MINOR_THEME More specific thematic category, when available.
SUBJECT Normalized subject or event label grouping related entries.
IMAGE_REFERENCE Image filename when applicable. Full paths or external references are not distributed.
MESSAGE Textual content of the news item, claim, reaction, comment, or verification-related material.
TYPE_MESSAGE Message type, for example FAKE, GENUINE, FAKECOMMENT, or related categories depending on the annotation scheme.
FAKE_CATEGORY Misinformation category when applicable, for example fabricated or misleading content.
LANGUAGE Language or dialect label, for example MSA, FRA, dialectal Arabic, Arabizi, English, or code-switched content.
PUBLICATION_DATE Publication date when available.
COMMENTS_NUMBER Number of comments associated with the source item when available.
SITE Source platform or website name, normalized where possible.
NUMBER_FOLLOWERS Follower count metadata when available.
NUMBER_OF_FORWARD Forward/share/repost count metadata when available.
COUNTRY Country metadata when available.
GENDER Gender metadata when available or inferable from the original source; UND indicates undetermined.
USER Anonymized user identifier. This field must not be used for re-identification.

Researchers should verify the meaning of each label before training or evaluating models, especially when grouping news items with their associated comments or reactions.

Recommended Usage Notes

When using BOUTEF for misinformation detection, researchers should clearly specify:

  • whether the task is performed at the message level, claim level, news-item level, or grouped news-plus-comments level;
  • how comments are linked to source items;
  • whether comments are used as input features, auxiliary context, or separate prediction targets;
  • whether temporal information is used, ignored, or used for temporal splitting;
  • whether metadata fields such as platform, follower count, publication date, or share count are included;
  • whether user-related fields are excluded from modeling to avoid profiling or privacy leakage;
  • how class imbalance is handled;
  • how multilingual, dialectal, and code-switched content is preprocessed.

For privacy-preserving research, it is recommended to exclude USER from model inputs unless there is a specific, ethically justified, and approved reason to use anonymized user-level grouping.

Ethical Considerations

BOUTEF contains sensitive social-media and misinformation-related content from North African contexts. Even after anonymisation, researchers must treat the dataset as controlled-access data.

Potential risks include:

  • re-identification of users or accounts;
  • amplification of harmful misinformation narratives;
  • misuse for surveillance or political targeting;
  • stigmatization of communities, languages, dialects, or regions;
  • overgeneralization from platform-specific or temporally bounded data.

Researchers should avoid presenting model results as definitive judgments about individuals, communities, or political groups. Dataset use should remain limited to responsible research, evaluation, and educational contexts.

Copyright, Platform Terms, and Third-Party Rights

Some materials in the dataset may originate from third-party online platforms, news organizations, fact-checking sources, or user-generated social-media content.

The dataset providers make no representations or warranties that the dataset is free from copyright, database rights, platform restrictions, or other third-party rights.

Researchers are solely responsible for ensuring that their use of the dataset complies with:

  • applicable laws;
  • institutional policies;
  • research ethics requirements;
  • platform terms of service;
  • copyright and database-rights obligations.

Publication and Citation

Researchers may publish aggregate results, analyses, statistics, trained models, or evaluation findings derived from the dataset, provided that such publication does not redistribute the dataset or reveal identifiable user information.

Any publication using the dataset must cite the associated dataset paper:

@misc{smaili2026boutefmultilingualcorpusfakenews,
      title={BOUTEF: A Multilingual Corpus for FakeNews in North Africa -- Language as a Weapon}, 
      author={Kamel Smaili and Yassine Toughrai and Amina Laggoun and David Langlois},
      year={2026},
      eprint={2606.00193},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.00193}, 
}

Controlled Research Use Data Agreement

1. Dataset

This agreement governs access to and use of the BOUTEF dataset.

The Dataset is a controlled-access research resource for the study of misinformation in North African social media. It contains misinformation narratives, genuine reference items, user-generated reactions, and debunking or verification-related material collected from online sources and social media platforms.

The Dataset includes multilingual and multidialectal content, including Modern Standard Arabic, Algerian dialect, Tunisian dialect, Moroccan dialect, Arabizi, French, English, and code-switched text.

By requesting, accessing, downloading, or using the Dataset, the user agrees to the terms below.

2. Permitted Use

The Dataset may be used only for non-commercial research and educational purposes.

Permitted uses include:

  • academic research;
  • benchmarking and evaluation;
  • reproducibility studies;
  • development of methods for misinformation detection;
  • multilingual and dialectal NLP research;
  • analysis of user reactions, debunking, and misinformation discourse.

The Dataset may not be used for commercial purposes without separate written permission from the Dataset providers.

3. Prohibited Use

The Researcher shall not use the Dataset for:

  • commercial product development or paid services;
  • advertising, marketing, or user profiling;
  • political targeting, persuasion, or influence operations;
  • surveillance or monitoring of individuals or communities;
  • law-enforcement profiling;
  • harassment, discrimination, or targeting of individuals or groups;
  • generation, amplification, or optimization of misinformation;
  • attempts to infer sensitive personal attributes of users;
  • any use that may cause harm to individuals, communities, or vulnerable groups.

4. Controlled Access and Redistribution

Access to the Dataset is granted on an individual basis.

The Researcher shall not:

  • redistribute the Dataset;
  • mirror or republish the Dataset;
  • upload the Dataset to public repositories;
  • sublicense, sell, or transfer the Dataset;
  • share the Dataset with unauthorized users;
  • expose the Dataset through public APIs, demos, or applications.

The Dataset may not be uploaded to third-party services unless those services provide equivalent access control, confidentiality, and data-protection safeguards.

The Dataset providers reserve the right to approve, deny, suspend, or terminate access at any time.

5. Research Team Access

The Researcher may provide access to students, research assistants, or collaborators working under their direct supervision only if each person has agreed to be bound by these terms.

The Researcher remains responsible for ensuring that all collaborators comply with this agreement.

Access must not be shared with individuals outside the approved research team.

6. Privacy and Non-Identification

The Researcher acknowledges that the Dataset may contain user-generated content, social-media reactions, and platform-derived metadata.

The Researcher shall not attempt to:

  • identify;
  • re-identify;
  • contact;
  • track;
  • profile;
  • deanonymize;

any individual, author, commenter, account, or social-media user represented in or associated with the Dataset.

The Researcher shall not combine the Dataset with external resources in a way that could reasonably enable re-identification.

The Researcher shall not publish raw examples, screenshots, links, metadata, or model outputs that could reveal the identity of users, accounts, or private individuals.

7. Copyright, Platform Terms, and Third-Party Rights

The Researcher acknowledges that some materials in the Dataset may originate from third-party online platforms, news organizations, fact-checking sources, or user-generated social-media content.

The Dataset providers make no representations or warranties that the Dataset is free from copyright, database rights, platform restrictions, or other third-party rights.

The Researcher is solely responsible for ensuring that their use of the Dataset complies with:

  • applicable laws;
  • institutional policies;
  • research ethics requirements;
  • platform terms of service;
  • copyright and database-rights obligations.

8. Publication and Citation

The Researcher may publish aggregate results, analyses, statistics, trained models, or evaluation findings derived from the Dataset, provided that such publication does not redistribute the Dataset or reveal identifiable user information.

Any publication using the Dataset must cite the associated dataset paper.
Downloads last month
21
Total file size:
76.6 MB

Paper for TeamSmart/BOUTEF