verifiable synthetic data

Synthetic data you can verify.

Plexoria builds tabular datasets for machine learning where the real data is legally locked away — with interpretable schemas, labelled hard cases, a benchmark in the box, and an Ed25519 signature on every bundle.

A trust brand for synthetic data.

Anyone can generate rows. The hard part is giving a buyer a reason to trust them. Every Plexoria dataset ships the same way — measured, documented, and signed — so you can check what you got instead of taking our word for it.

01 Interpretable

Named, meaningful columns — not anonymised PCA components. Do real feature engineering and understand the data.

02 Labelled hard cases

Fraud is split into named typologies, including covert ones that hug the legitimate distribution — the cases detectors actually miss.

03 Proven, not asserted

A reproducible baseline benchmark ships in every bundle. The quality is a number you can re-run, not a claim.

04 Signed provenance

SHA-256 manifest + Ed25519 signature, with the key fingerprint published out-of-band. Prove the bundle is intact and from us.

Synthetic card-fraud dataset.

50,000 fully-synthetic card transactions at a realistic 0.5% fraud rate, across five labelled typologies. License-clean and GDPR-safe — no real transaction, card, or person is involved. A baseline benchmark proves the signal is learnable and that the covert typologies are genuinely hard.

0.945
ROC-AUC · global
0.72
counterfeit_present
0.66
friendly_fraud
1.00
privacy score
TypologyKindWhat it is
card_testingovertMicro-amount bursts probing a stolen number.
cnp_ecommerceovertStolen card used online, elevated amount.
account_takeoverovertHigh-value cash-out, foreign, at night.
counterfeit_presenthardCloned card used in-store — looks in-person.
friendly_fraudhardA genuine purchase later disputed — near-legit.

Generate · validate · learn.

New to synthetic data? The Learn guide explains the whole pipeline from scratch — the vocabulary, how a copula imitates a table, the three validation tests, and how a model learns from the result. No prerequisites.

Read the guide →