NaijaScamBench (v0)

A small, labelled benchmark of Nigerian scam and legitimate messages (SMS and WhatsApp style) in English, Nigerian Pidgin, Yoruba, Hausa and Igbo. It was built to evaluate Xvimo, an open-source scam checker for Nigeria that runs on NVIDIA Nemotron models.

Most public spam datasets are Western SMS. Nigerian fraud looks different: fake job interviews at a physical venue, BVN and ATM "block" alerts, Ponzi "investment" schemes, fake loans, giveaway and romance lures, often written in Pidgin or a local language.

Contents

Field	Description
id	Message ID (matches the Xvimo benchmark logs)
text	The message. Private details are replaced by placeholders such as [NAME], [PHONE], [ACCOUNT], [AMOUNT], [CODE]
label	scam or legit
scam_type	ponzi, phishing_link, fake_loan, advance_fee, impersonation, fake_giveaway, romance, fake_job, or none
language	english, pidgin, yoruba, hausa, igbo, mixed
channel	Where such a message usually arrives (SMS, WhatsApp, …)
has_link	Whether the message contains a link
source	constructed or UCI SMS Spam Collection (CC BY 4.0)

180 messages: 49 scam, 131 legit.

Language	Scam	Legit
English	37	118
Pidgin	5	7
Yoruba	2	2
Hausa	2	2
Igbo	2	2
Mixed	1	0

Sources

• Constructed (135): written by the Xvimo team to reflect common Nigerian scam and everyday message patterns. These contain no real people's data.
• UCI SMS Spam Collection (45 legit "ham" messages): Almeida, T. & Hidalgo, J. (2011), UCI Machine Learning Repository, CC BY 4.0. They are included as casual everyday chat.

The full Xvimo benchmark uses 330 messages. The other 150 come from sources whose licence does not allow redistribution, so they are not published here:

• 120 real Nigerian scam SMS (job and financial scams) from a public repository built on Data Science Nigeria data;
• 30 everyday Nigerian messages.

Leaderboard numbers in the Xvimo README refer to the full 330.

Limitations (please read)

• Small. Treat this as a starting point, not a definitive test.
• Local languages are very thin: 4 messages each for Yoruba, Hausa and Igbo. A score of 100% on 4 messages proves very little. These messages have not yet been reviewed by native speakers.
• Placeholders like [ACCOUNT] make some legitimate bank alerts look unusual to models, which can inflate false alarms.
• Constructed messages were written by the same team that built Xvimo. Scores on them may be optimistic.

Contributing

Native-speaker corrections and new real-world examples (with personal details removed) are very welcome. Open a discussion on this dataset page or an issue on the Xvimo GitHub repository.

Citation

@misc{naijascambench2026,
  title  = {NaijaScamBench: Nigerian scam and legitimate messages for evaluating scam checkers},
  author = {Xvimo team},
  year   = {2026},
  url    = {https://huggingface.co/datasets/<your-username>/NaijaScamBench}
}