NaijaScam v0.1

The first open dataset of Nigerian scam messages in English, Nigerian Pidgin, Yorùbá, Igbo and Hausa. 1,821 labelled SMS / WhatsApp / DM / email-style messages (982 scam, 839 legitimate) covering the nine scam patterns that hit Nigerians hardest, plus eight kinds of genuine messages that look similar (bank alerts, OTPs, security advisories…). Built for Shora (Ṣọ́ra), an open-source scam shield.

⚠️ Honesty first: this is synthetic-but-grounded data. Messages were generated by an LLM (openai/gpt-oss-120b via Groq) with prompts seeded by patterns described in public warnings from the CBN, banks, EFCC, FCCPC, NIS/INEC, fact-checkers and news reports, plus 36 hand-written seed messages. No real person's messages, names, phone numbers or account numbers are included (all are fake or masked). It is a starting point, not a substitute for real-world data — see Limitations.

Quick start

from datasets import load_dataset
ds = load_dataset("Nihilitybot666/naijascam")          # train / validation / test
challenge = load_dataset("Nihilitybot666/naijascam", "challenge")["test"]  # 40 hand-written OOD messages

Fields

field	type	description
id	string	stable id (ns-00042, ch-007 for the challenge set)
text	string	the message
label	scam / legit	binary label
scam_type	string	one of 9 scam types or 8 legit types (below)
language	en pcm yo ig ha	main language (code-mixing with English is normal and kept)
red_flags	list[string]	short English phrases naming the warning signs (empty for legit)
source	synthetic / seed	LLM-generated vs hand-written seed
generator	string	groq/openai/gpt-oss-120b or hand-written

Taxonomy and counts

scam_type	label	en	pcm	yo	ig	ha	total
bank_impersonation_otp — bank/BVN/NIN impersonation & OTP theft	scam	50	27	13	11	13	114
fake_credit_alert — fake alerts, overpayment refunds, reversal fraud	scam	52	24	12	11	12	111
fake_job_visa_scholarship	scam	47	27	12	12	12	110
fake_vendor_pos — online vendors, fake delivery fees, POS tricks	scam	47	26	11	12	12	108
giveaway_grant — fake FG/CBN grants, brand giveaways	scam	50	22	12	13	13	110
loan_app_extortion — illegal lender threats, upfront-fee loans	scam	52	25	12	9	12	110
ponzi_crypto — CBEX-style "AI trading", fixed daily returns	scam	46	27	13	12	13	111
romance_yahoo — romance / customs-package / emergency money	scam	38	35	12	12	12	109
sim_swap — fake telco/NCC SIM upgrade, PUK requests	scam	42	21	12	12	12	99
legit_bank_alert	legit	46	25	12	12	12	107
legit_business_vendor	legit	42	29	12	11	12	106
legit_delivery	legit	38	30	12	11	12	103
legit_family_chat	legit	37	32	11	13	13	106
legit_otp	legit	40	33	12	12	12	109
legit_school_remita	legit	42	25	11	12	12	102
legit_security_advisory — genuine "we will never ask for your PIN" notices (hard negatives)	legit	35	32	12	12	11	102
legit_telco_utility	legit	43	25	12	12	12	104
total		747	465	203	199	207	1,821

Splits (stratified by label × language, seed 42): train 1,274 · validation 182 · test 365. The test split is held out (no text overlap with train/validation, enforced by a unit test) and is also shipped as human-readable test.csv. The challenge config holds 40 hand-written messages in a different, messier style, used as an out-of-distribution check.

How it was built

1. Grounding. Tavily web searches collected public descriptions of each scam pattern (CBN BVN-fraud statements, POS fake-alert reporting by HumAngle and others, EFCC/CBEX coverage by France 24 and Al Jazeera, FCCPC loan-app guidance, NIS/INEC fake-recruitment warnings, haltfake.org and Africa Check on fake grants, Paga's giveaway alert). Source list: data/seeds/tavily_seeds.json.
2. Generation. 153 prompts (17 types × 5 languages, 12 messages each) asked for realistic messages with varied channels, banks/fintechs (GTBank, Access, Zenith, UBA, OPay, Moniepoint, Palmpay, Kuda…), cities, naira amounts, masked accounts and fake links. A third of each batch was requested as subtle scams or hard-negative legit messages. Raw output: data/raw/naijascam_raw_v0.1.jsonl (1,834 rows) + 36 seeds.
3. Cleaning & quality filters (scripts/build_dataset.py): stripped channel prefixes ("SMS:"), randomised repeated account masks, dropped legit messages that ask for OTP/PIN/BVN or push unknown short links (35), scams without red flags (13), length outliers; relabelled messages written in plain English despite a Pidgin/Yoruba/Igbo prompt (125 → en); removed exact and near duplicates (char 3–5-gram cosine > 0.85).

Benchmark (NaijaScamBench v0.1)

Model	Test acc.	Test macro-F1	Challenge acc.	Challenge scam recall
Rules only	72.0	72.0	60.0	—
TF-IDF + LogReg	97.8	97.8	75.0	50% (10/20)
afro-xlmr-mini fine-tuned	94.2	94.2	67.5	35% (7/20)
gpt-oss-120b zero-shot (partial)	—	—	—	95% (19/20)

Full table with per-language F1: benchmark/RESULTS.md in the GitHub repo.

Limitations

• Synthetic. Generated text is cleaner and more "on-topic" than real scams; in-distribution test scores are inflated. The hand-written challenge set shows a large drop for small models.
• Language quality. Yorùbá, Igbo and Hausa text was LLM-written and only spot-checked; tone marks and idioms may be imperfect. Native-speaker review is the top contribution we need.
• Single generator (one LLM family), so stylistic artefacts are shared across splits.
• red_flags are LLM-written English phrases, not span annotations.
• Scam tactics change fast (new grant names, new "AI" platforms); v0.1 reflects 2025–26 patterns.

Ethical use

For building defences, research and education. Do not use it to craft scams. Contributions of real messages must have all personal data removed (see CONTRIBUTING.md in the repo).

Citation

@misc{kpeale2026naijascam,
  title  = {NaijaScam v0.1: A Multilingual Nigerian Scam Message Dataset and Benchmark},
  author = {Kpeale, Akara},
  year   = {2026},
  url    = {https://github.com/kpealeakara-rgb/shora}
}