kadian-arch/stop-arnaque-237

Stop Arnaque 237: Cameroon Scam Reports Dataset

Scam and genuine messages from Cameroon, labelled scam or not_scam, in English, French and Cameroonian Pidgin. Real scam messages sit next to the genuine MTN and Orange messages they imitate (money alerts, promos, security tips), so a detector can learn to tell them apart instead of flagging every money message.

Version 1.0. Source, documentation and pipeline: https://github.com/kadian-arch/stop-arnaque-237. DOI of this version: https://doi.org/10.5281/zenodo.23287618 (all versions: https://doi.org/10.5281/zenodo.23287617)

Load it

from datasets import load_dataset

msgs = load_dataset("kadian-arch/stop-arnaque-237")                      # train and test splits
reports = load_dataset("kadian-arch/stop-arnaque-237", "reports")        # full scam reports behind the messages

What is in it

Config	Rows	One row is
messages (default)	267 (218 train / 49 test)	one message, scam or genuine, labelled
reports	60	a scam reported by a person in Cameroon: what happened, what was asked, money lost, region
public_alerts	160	a scam documented by fact-checkers, institutions or news sites, or shown in a public social media post
genuine_messages	180	a real message (not a scam) with its sender, operator and kind
scam_numbers	39	a scammer number (pseudonymized) and how often it appears

messages columns: id, text, label (scam / not_scam), scam_types, multi_scam, language (en, fr, pidgin, mixed, unknown), channel, text_type (verbatim, or retold by the person), origin (report, public_alert, social_post, contributed) and split.

The train/test split is keyed on the wording: near-identical messages always land in the same split, so a model is never tested on a copy of what it learned from. At most 3 messages with the same wording are included.

Every column and allowed value: data_dictionary.csv. Scam type codes: TAXONOMY.md in the GitHub repository.

Privacy

Nothing personal is released. Phone numbers become the operator prefix plus a 6-letter keyed code ([PHONE:676-kqbwmx]), so repeat scammers link up but no number can be recovered or dialled. Names, account numbers, codes and transaction ids are removed, links are made unclickable and screenshots are never published. Public social media posts keep only the scam message and a summary written by the project. Details: DATASHEET.md and PRIVACY.md in the repository.

Intended use

Research on scams and social engineering in Cameroon and Central Africa, training and evaluating scam detectors, awareness material and journalism. Not for identifying or contacting anyone. A number in this dataset may be spoofed, stolen or recycled.

Limitations

Small and growing. Early reports come mostly from the South West region and English-speaking groups. Amounts are self-reported. Public alerts over-represent fake job adverts, because that is what institutions publish about. Some report texts are the person's retelling rather than the original message (text_type = retold).

Citation

Kum, D. A. (2026). Stop Arnaque 237: Cameroon Scam Reports Dataset (Version 1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23287618

@dataset{kum_2026_stop_arnaque_237,
  author    = {Kum, Donalsien Akwo},
  title     = {Stop Arnaque 237: Cameroon Scam Reports Dataset},
  year      = {2026},
  version   = {1.0},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23287618},
  url       = {https://doi.org/10.5281/zenodo.23287618}
}

Licence: data CC BY 4.0, code MIT. Contact and removal requests: groundtruth.cm@gmail.com