AfriScam

A Dataset of Scam and Fraud Messages from Kenya and Nigeria.

A dataset of scam and fraud messages from African countries, labelled with a shared taxonomy of scam types, persuasion tactics, claimed identities, requested actions and verification status. It supports research, evaluation of scam-detection models, and public awareness work. Work in progress.

Contents

Path	Description
data/kenya/af-scam-ke-v0.1.csv	Kenya observations: 60 records, 10 real-world first, then 50 synthetic
data/nigeria/af-scam-ng-screenshots-v0.1.csv	Nigeria observations from public screenshots: 53 real-world records (pending review)
data/nigeria/af-scam-ng-contributor-R-F8JM-W6U-v0.1.csv	Nigeria observations from a contributor: 10 real-world records (pending review)
taxonomy/scam_taxonomy.csv	Category → scam type → subtype hierarchy (68 subtypes) with working definitions
taxonomy/tactics.csv	Controlled vocabulary of 22 persuasion tactics
docs/data_dictionary.md	Field definitions and allowed values
docs/annotation_guidelines.md	Annotation rules and open taxonomy questions
docs/verification_protocol.md	How fraud status is established and recorded
docs/privacy_and_consent.md	Redaction, consent and release rules
examples/	Usage examples

Record origin

Each record is either synthetic (written to illustrate a taxonomy subtype) or real_world (a message contributed by the person who received it). The two are never mixed: synthetic records are never marked verified. Filter on record_origin when evaluating models.

Real-world records may have a screenshot_file pointing to the original screenshot. Screenshots are unredacted and are not published in this repository.

Format

• UTF-8 CSV, one row per message.
• data/afriscam-v0.1.jsonl combines all the CSVs into one file, one JSON object per line. List columns are real lists, and source_file names the CSV each record came from. Rebuild it with python3 scripts/combine_data.py. The Hugging Face viewer shows this file.
• tactics, requested_action, risk_indicators and language_tags are JSON arrays, e.g. ["authority_claim", "urgency_pressure"]. An empty cell means not yet labelled.
• language_tags use ISO 639-1 codes (en, sw). Sheng has no ISO code and is tagged sw, with a note in review_notes.
• Personal and financial details are replaced with placeholders: [PHONE], [LINK], [EMAIL], [PAYBILL], [ACCOUNT], [AMOUNT], [ORGANIZATION], [NAME], [ADDRESS], [CODE].
• Spelling and grammar in message text are kept exactly as received, because errors are themselves scam signals.

Quick start

Load the dataset from Hugging Face:

from datasets import load_dataset

ds = load_dataset("Afrieval/AfriScam", split="train")
df = ds.to_pandas()

real = df[df.record_origin == "real_world"]
print(real.explode("tactics").tactics.value_counts())

Or read the file from a clone of this repository:

import pandas as pd

df = pd.read_json("data/afriscam-v0.1.jsonl", lines=True)

tactics, requested_action, risk_indicators and language_tags load as lists.

Status

Version 0.1 is a draft. All records are pending_review.

• Working files: each CSV is exported from an Excel workbook used for annotation (AF-SCAM-KE_Standardized_v0.1.xlsx for Kenya, and the matching .xlsx next to each Nigeria CSV). The workbooks also hold verification events and review notes, and are not published.
• Nigeria duplicates: one public screenshot record duplicated a contributor's firsthand submission (AF-SCAM-NG-REAL-0071). The contribution is kept and the screenshot copy removed; its ID AF-SCAM-NG-REAL-0010 is retired.
• Out of scope: eight screenshot records from outside Africa were removed and their IDs retired: AF-SCAM-NG-REAL-0005, 0058 and 0053 (US, US, Singapore), 0033, 0046 and 0047 (Australia), 0057 (Philippines) and 0060 (India). Subtypes TAX-058 and TAX-065 now have no records but stay in the taxonomy.
• Negative examples: four screenshot records judged legitimate on review are kept with fraud_status = not_scam (AF-SCAM-NG-REAL-0023, 0029, 0034, 0048). Filter them out when you need scam messages only.
• Taxonomy: 12 subtypes (TAX-057 to TAX-068) and one tactic (religious_appeal) were added from the Nigeria data, including BVN update phishing, fake bank credit alerts and fake password change alerts.

See CHANGELOG.md for changes.

Authors

Tosin Amuda, Chanice Wanjiku Irungu

Citation

@misc{amuda2026afriscam,
  title        = {AfriScam: A Dataset of Scam and Fraud Messages from Kenya and Nigeria},
  author       = {Amuda, Tosin and Irungu, Chanice Wanjiku},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Afrieval/AfriScam}}
}

License

See LICENSE.