abidlabs/synthetic-fraud-100

Synthetic fraud detection dataset (100 samples)

A small synthetic tabular dataset for fraud-detection demos, pipeline tests, and teaching. 100 card transactions, 50 fraud / 50 legit, readable features, and a stratified 80/20 train/test split column.

Files

• synthetic-fraud-100.csv — 100 rows, 11 columns.

Schema

Column	Type	Description
transaction_id	text	Synthetic id, T0001–T0100
amount	float	Transaction amount in USD
hour	int	Hour of day, 0–23
merchant_category	category	retail, restaurants, fuel, clothing, online_services, electronics, travel, jewelry
channel	category	in_person or online
distance_from_home_km	float	Distance from cardholder's home; 0.0 for online transactions (does not apply)
is_foreign	0/1	Foreign country / cross-border transaction
new_device	0/1	First seen device
txns_last_24h	int	Number of transactions on the card in the past 24h (velocity)
is_fraud	0/1	Label: 1 = fraud
split	category	train (80 rows) or test (20 rows), stratified by label

Splits

• train: 80 rows — 40 fraud / 40 legit
• test: 20 rows — 10 fraud / 10 legit

Use the split column; the split is stratified, so a model trained on split == "train" can be evaluated fairly on split == "test".

Design notes

Three fraud patterns are mixed together so the task is not trivially separable:

1. Card-testing bursts — tiny amounts ($1–10), online, new device, high velocity (9–15 transactions/day).
2. Stolen-card big spends — jewelry/travel/electronics ($600–4,100), far from home (400–3,100 km), often foreign, often odd hours.
3. Mild frauds (~10 rows) — amounts and times look normal; only one signal is off (e.g., an unusually far-from-home grocery→retail run).

The legit side contains hard negatives on purpose: large travel/electronics purchases, some foreign, some on new devices — so amount alone or a single flag does not solve the task.

Limitations

• Fully synthetic and hand-authored; it does not reflect any real population of transactions and should never be used to draw conclusions about fraud in the real world.
• Balanced 50/50 labels, unlike production fraud data (typically <1% fraud). Models trained on this will not transfer to realistic class imbalance.
• 100 rows is far too small to train anything general; use it for demos, unit tests of training pipelines, and teaching.

Intended use

Quick demos (e.g., a Gradio fraud-classifier app), integration tests for training/eval pipelines, and classroom examples where a small, human-readable tabular dataset with a known label balance is more useful than a large anonymous one.