abidlabs/synthetic-fraud-demo

Synthetic Fraud Detection Demo (10 samples)

Ten hand-written synthetic card-transaction records for smoke-testing a fraud-detection pipeline: loading, schema validation, feature engineering, and a demo UI.

This is not training or evaluation data. Ten rows cannot support a model, and the class balance here (4 fraud / 6 legitimate, 40%) is nothing like real card fraud prevalence, which is roughly 0.1–0.5%.

Loading

from datasets import load_dataset

ds = load_dataset("abidlabs/synthetic-fraud-demo", split="train")
print(ds.to_pandas())

Schema

column	type	description
transaction_id	string	Unique transaction identifier
timestamp_utc	string	ISO-8601 authorization timestamp (UTC)
amount_usd	float	Transaction amount in US dollars
merchant_category	string	Merchant category code group
payment_method	string	card_present or card_not_present
card_country	string	ISO-3166 alpha-2 country of the card issuer
ip_country	string	ISO-3166 alpha-2 country of the originating IP
distance_from_home_km	float	Distance between the transaction and the cardholder's home
txn_count_24h	int	Number of transactions on the card in the preceding 24 hours
account_age_days	int	Age of the account at authorization time
is_new_merchant	bool	Whether this is the first transaction with the merchant
is_fraud	int	Label: 1 = confirmed fraud, 0 = legitimate

How the label was assigned

The label is deliberately correlated with the features, so the set is learnable: fraudulent rows are card-not-present, have a card/IP country mismatch, are far from home, show high 24-hour velocity, belong to young accounts, and go to new merchants. Legitimate rows are mostly card-present, local, and on established accounts. A model trained on this will separate the two easily — which is the point for a pipeline test and useless as a signal-quality claim.

Notes

• All records are synthetic. No real cardholder data is involved.
• Timestamps span 2026-09-14 to 2026-09-16, so the rows have a usable time ordering.