Turkish retail banking data your test environment can hold without a KVKK conversation.
Customers, accounts, cards, transactions, and the free-text surfaces where
personal data actually hides: complaints, KYC notes, and support transcripts.
Mint it yourself · The evaluation set · What is in here · Türkçe
This repository contains no engine code. It is declarations and data: record shapes, lexicons, templates, and recipes. The engine that reads them lives in mintmark and is pinned here by a version range with a closed upper bound.
Turkish banks and fintechs cannot move production data into test, evaluation, or AI pilot environments without KVKK exposure, and their vendors cannot reasonably ask them to. This pack declares the data those environments need, and the engine mints it: deterministic, span-labeled, and sealed by a manifest anyone can check.
Release 0.2.0. Its reference datasets are attached to
v0.2.0,
each carrying its own manifest and checksums. packcheck passes against the
pinned core, the test suite passes, and the evaluation recipe meets every one of
its eighteen coverage targets.
Important
What this pack is not. It is not anonymization of your banking data; it ingests none. It is not a compliance guarantee and not a legal safe harbor. The anomaly labels are detector-side test data and this repository documents no evasion guidance of any kind. Generated phone numbers can coincide with assigned ones, because the Turkish numbering plan reserves no fictional range. This data is for testing systems. It is never for contacting anyone. What this does and does not mean under Turkish data protection law is set out in docs/kvkk.md.
View the accessible SVG source
| In here | Not in here |
|---|---|
| Seven record types, three of which are free text | Any engine code. The only Python is under tests/ and imports the public API alone |
| 26 invented bank names, scanned against a real-institution register in CI | Any real bank, merchant, or person. A collision fails the build |
| Two recipes, one of which is a labeled evaluation set | A dataset in git. Only bounded samples, capped at 50 records per type |
| Every identifier checksum-invalid by default | A corporate customer record type. This pack is retail, which is why tax numbers reach it through documents |
uv tool install mintmark
git clone https://github.com/lokomotifai/mintmark-banking
cd mintmark-banking
mintmark packcheck .
mintmark mint --pack . --recipe retail-baseline --seed 20260901 --out ./run
mintmark verify ./runOffline after dependency bootstrap. packcheck validates the declarations under
the strict loader, mini-mints them, and runs the invariant and denylist scans.
Want a look first, without minting anything? samples/ carries fifty
records of every type, and a test regenerates them from a fixed seed and compares
by digest, so a sample that drifted from the declarations fails the build rather
than quietly misrepresenting them.
One complaint body, as emitted:
Sayin yetkili, Meryem Şimşek adima kayitli hesabimla ilgili bir
sorun yasiyorum. kredi konusunda, sube yogunlugu basliginda birkac
kez basvurmama ragmen cozum alamadim. Kimlik numaram 45797577307,
hesabim TR439999902711375762415596. Adresim Fatih Mahallesi,
ulasilabilecegim numara +90 592 647 78 19. Konunun
degerlendirilmesini ve tarafima geri donus yapilmasini rica
ederim.
That is the first record in samples/complaint_ticket.jsonl,
not an illustration written for the README. A test compares the two, so the
example cannot drift from what the pack actually emits.
Every value in it is synthetic, and the identity number fails its own check digit rule. The sidecar for that document records a span for each of them.
The pii-eval recipe is the reason this pack exists in the shape it does. The
hushmark-tr model card asks adopters to evaluate a detector on representative
data before production use; this is that data for Turkish banking.
It declares a coverage target for every label and meets all eighteen:
| Label group | Target | Achieved |
|---|---|---|
| PERSON, ADDRESS, ORG, DOB | 300 each | 1178 to 2275 |
| HEALTH, RELIGION, ETHNICITY, POLITICAL | 300 each | 483 to 512 |
| SEXUAL_LIFE, CRIMINAL, BIOMETRIC_REF, UNION | 300 each | 486 to 509 |
| TCKN, IBAN, PAN, PHONE, EMAIL | 500 each | 768 to 2095 |
| VKN | 500 | 1268 |
That last row is worth a note. Every record type in this pack is retail, so there is no corporate customer to carry a tax number in a field. VKN reaches the data through document templates instead, in the phrasings where a tax number genuinely appears in Turkish banking text: a disputed corporate mandate, a sole proprietor confirming one during KYC, an invoice payment naming a merchant's.
The arithmetic behind the special-category rows is worth stating too, because it is not obvious. Eight labels at 300 spans each is 2400 injections across 2000 documents, which is more than one per document. The baseline rate of 0.02 would have produced about forty. The evaluation recipe therefore uses a separate template family at a rate of one, with two special-category slots per document and the labels spread evenly across the set. Baseline and evaluation templates are separate sets rather than one set with a knob, so that a recipe cannot half mix them.
An identifier inside a document body is a fresh draw. {id:TCKN}, {id:IBAN},
{id:PHONE}, and the person named by {entity:PERSON} are drawn independently
of the customer or account record the document is attached to. The spans remain
correct, but this pack does not support tests that require document and record
identities to agree.
| Recipe | Shape | For |
|---|---|---|
| retail-baseline | 10 000 customers, about 18 000 accounts, 9 000 cards, 250 000 transactions, and 2 800 documents | Filling a test environment with something that behaves like a portfolio |
| pii-eval | 2 000 documents, every label above its target | Measuring a detector's recall and precision on Turkish banking text |
Every transaction carries anomaly_kind and is_anomaly, and the two never
disagree. But the four kinds are per-row labels drawn at declared rates, not
genuine temporal structures. A real burst is many transactions clustered in
time on one account; here it is a label.
That is a limit of the pack contract rather than an oversight: each field is drawn from an independent stream, so a pack cannot declare a pattern that correlates rows. Genuine temporal shapes need a core change, and it is recorded as one. Use the baseline recipe to check that your pipeline carries labels through correctly. Do not use it to measure whether a detector finds real bursts.
Inherited from the core, and re-checked on the artifacts by verify rather than
promised:
- TCKN and VKN are computed correctly and then corrupted by a nonzero offset, so they fail the exact rule a validator applies.
- IBAN carries the bank code
99999, verified absent from the published payment systems participant register. Even a validator-mode IBAN names no real institution. - PAN begins with
9, a major industry identifier no commercial card network uses, and is emitted masked. - EMAIL sits only under names reserved by RFC 2606 and RFC 6761, which nobody can register.
Every reference dataset this repository publishes is minted with the safe policy. A test asserts that for each of them.
The 26 bank names here follow a place-and-nature pattern rather than a person-name one, because person-derived institution names collide with real entities more often and can resemble a real person's business.
Every one is scanned in required CI against a denylist built from the published participant register of licensed banks. A collision fails the build and names both sides of it. A name that collides is removed, not defended.
The check runs over lexicons, templates, and minted output, so a real name cannot enter through a template literal either.
pack.yaml identity, the core pin, the allowed identifier policies
fields/ one file per record type, in generation order
recipes/ retail-baseline and pii-eval
templates/ baseline sets, and the separate evaluation sets
lexicons/ invented banks, products, counterparties, and the denylist
samples/ fifty records per type, regenerated from a fixed seed
vendor/ the core wheel required CI runs against, recorded by checksum
tests/ the conformance suite
docs/ the reference dataset record and engineering notes
uv sync
uv run mintmark packcheck .
uv run pytest
uv run python tools/mdlint.py .All of it runs offline against the vendored core wheel. A separated, network-labeled workflow confirms weekly that the vendored artifact still matches the core repository at the pinned tag, because the offline check is only as good as the artifact it runs against.
Version 0.2.0 is released. Its two reference datasets are attached to
v0.2.0;
they were minted with the safe identifier policy at the seeds declared in
docs/reference-datasets.json. The engine is on PyPI as
mintmark.
The seeds are settled deliberately. A changed seed silently invalidates a published manifest, so it is never the fix for a coverage miss; the templates or the document mix change instead.
Under semantic versioning, the public surface here is pack.yaml, the field
declarations, the recipes, the template sets, the lexicons, and the bytes a fixed
seed produces. After the first tagged release, adding an entry to a lexicon
changes the draw for every subsequent index, which makes it a major bump rather
than a minor one.
Contributions under the Developer Certificate of Origin 1.1, no contributor license agreement. See CONTRIBUTING.md, which explains why there is no engine code here and what to do when a declaration cannot express what the pack needs. GOVERNANCE.md sets out what this repository decides and what it does not. SECURITY.md covers the private reporting route and what counts as a vulnerability in a repository that executes nothing.
README.md is canonical and README.tr.md is a full mirror.
Apache-2.0. See LICENSE and NOTICE. The license grants no right to the Mintmark name or logo; see TRADEMARKS.md.
Reference datasets are licensed CC BY 4.0: use them for anything, including
commercially, and credit the source. Every dataset carries its own credit line in
MINTMARK.json and mintmark verify prints it, so nothing has to be assembled by
hand. See LICENSE-DATASETS.md.
Part of the Mintmark family: the engine · insurance · human resources
