Turkish workforce and payroll data, including the two surfaces most people would rather not model.
Employees with their position history, leave and payroll,
and the free text where health and criminal record mentions actually appear.
Mint it yourself · The two boundaries · What the contract decided · Türkçe
This repository contains no engine code. It is declarations and data. The engine that reads them lives in mintmark and is pinned here by a version range with a closed upper bound.
Every Turkish company holds employee data, and almost none of it can move into a test environment. Payroll carries bank details. Leave records carry sick days. Recruiting carries criminal record checks. This pack declares that data, and the engine mints it: deterministic, span-labeled, and sealed by a manifest.
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 HR system; it ingests none. It is not a compliance guarantee and not a legal safe harbor. It is not clinical data: health stays at category granularity by design. It makes no accusation about anybody, synthetic or otherwise: the criminal record surface records that a document was requested, never what it contained. Payroll amounts model no tax code and no minimum wage. 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 tools/ |
| Payroll with gross, net, an IBAN and a labeled anomaly kind | Any tax calculation. Net is a draw, not a deduction |
| A criminal record surface, reached through reference checks | Any statement about what a record contains. See below |
| 26 invented employer names, scanned against a real-company list | Any real company. The scan runs on every mint in CI |
Most packs in this family hold one sensitive surface. This one holds two, and both are held by a control rather than by review, because review gets tired and a list does not.
Health reaches HR text in two ordinary ways: a leave request that gives a reason, and a performance conversation that mentions a working arrangement. Both are in here. A condition class and nothing more: no diagnosis, no clinical finding, no treatment, no medication, no prognosis.
Every health span draws from the core's curated condition-class descriptors, and
lexicons/clinical_denied_tr.txt lists the vocabulary this pack refuses. A test
scans every rendered document against it. The failure mode is silent, which is
why the control exists: a template that drifts into clinical detail still renders,
still labels, and still passes every other check.
A reference check is a process. This pack's documents record that a criminal record document was requested, received, or is pending. They never record what it contains.
Text asserting that a person committed an offence, was convicted, or is under
investigation is an accusation about a person, and a synthetic person is still a
person-shaped claim. A model trained on text that convicts synthetic people learns
to produce that text about real ones. So
lexicons/accusatory_denied_tr.txt lists 43 phrases this pack refuses,
covering conviction, allegation, named offence classes, and workplace conduct
framed as an accusation. Every rendered document is scanned against it, and a test
plants a denied phrase to prove the list still catches one.
None of these is an oversight, and all three are the kind of thing a reader will otherwise assume is a bug.
A leave type carries no label, including the sick-leave code. leave_record.type
is a structured enum, and the last of its values is raporlu. Labeling it HEALTH
would teach a detector that a seven-character enum value in a database column is a
health disclosure, and every evaluation run against the dataset would then be
scoring detectors on whether they flag codes. The health signal worth detecting
lives in the document bodies, where a condition class appears in running Turkish
text and a span points at it.
A recruiter note points at no employee. Its subject is a candidate, and a candidate is by definition not yet an employee. A parent reference here would encode the opposite claim into every dataset the pack produces. The type is an orphan in the record graph, which is the honest shape of the data.
The manager name corresponds to no employee record. The pack contract has no
self-reference, so employee.manager_name is a surname lexicon draw. That is a
real modeling loss for anyone testing org-chart traversal, and it is stated here
rather than discovered after ingest.
uv tool install mintmark
git clone https://github.com/lokomotifai/mintmark-hr
cd mintmark-hr
mintmark packcheck .
mintmark mint --pack . --recipe workforce-baseline --seed 20261101 --out ./run
mintmark verify ./runOne recruiter note, as emitted:
Aday degerlendirme notu. Aday Ebru Özcan, basvurdugu pozisyon
icin teknik_degerlendirme asamasinda degerlendirildi. Iletisim +90
592 723 56 80, eposta kullanici7917.9145@example.org. Onceki
isvereni Deniz Kum Insaat. Teknik yetkinlik beklentiyi karsiliyor
olarak notlandi. Surec sonraki asamaya gecti.
That is the first record in samples/recruiter_note.jsonl,
not an illustration written for the README. A test compares the two.
pii-eval declares a coverage target for every label and meets all eighteen.
| Label group | Target | Achieved |
|---|---|---|
| PERSON, ADDRESS, ORG, DOB | 300 each | 1645 to 3000 |
| The eight special categories | 300 each | 714 to 776 |
| TCKN, VKN, IBAN, PAN, PHONE, EMAIL | 500 each | 806 to 3000 |
Eight special-category labels at 300 spans each is 2400 injections. The baseline recipe runs its special rate at 0.06, which across baseline document volume produces nowhere near that, and raising the baseline rate to reach the target would misrepresent how often special categories actually appear in HR text. So the evaluation twins are separate record types with their own template family at rate one, two special slots each, spread evenly across the labels. Three document families rather than the insurance pack's two makes the arithmetic comfortable here.
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 employee 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 |
|---|---|---|
| workforce-baseline | 6 000 employees, 11 000 position rows, 24 000 leave records, 72 000 payroll entries, and 11 000 documents | Filling a test environment with something that behaves like a workforce |
| pii-eval | 3 000 documents, every label above its target | Measuring a detector on Turkish HR text |
Every payroll entry carries anomaly_kind and is_anomaly, and the two never
disagree. But the kinds are per-row labels drawn at declared rates, not genuine
temporal or cross-record structures. A real payroll anomaly is a month that
breaks a pattern across an employee's other months; here it is a label on one row.
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. Use the baseline recipe to check that your pipeline carries labels through correctly. Do not use it to measure whether a detector finds real patterns.
Worth telling, because it is what a control is for and because the fix changed the rule rather than the symptom.
This pack invents employer names, and neither the core's bank list nor the insurance pack's insurer list was built to catch a collision with an industrial holding. So a compilation of 51 widely known Turkish corporates went into the pack denylist, each name reduced to its distinctive part.
That reduction turned Dogan Holding into dogan and Yildiz Holding into
yildiz. Both are among the most common surnames in Turkey, and both are drawn by
the core's own surname lexicon. The list fired four times on a single baseline
mint, every hit a synthetic employee rather than a real company.
A denylist that flags the pack's own data is a denylist somebody switches off, so
the rule changed: a one-word core is now checked against the core's given-name,
surname, province, and street lexicons, and a name whose distinctive part is an
ordinary Turkish word keeps the whole company name. The same rule is why
Toros Tarim is listed as two words, so an invented Toros Lojistik does not
collide while the real fertilizer producer is still covered.
One limitation is worth stating: the exchange publishes the authoritative listing of traded companies and it was not available as a machine-readable public endpoint from here, so the list used is a compilation rather than the register. A manual read of the exchange's own list belongs in the release checklist. The full record is in docs/normative-verification.md.
pack.yaml identity, the core pin, the allowed identifier policies
fields/ one file per record type, in generation order
recipes/ workforce-baseline and pii-eval
templates/ baseline sets, and the separate evaluation sets
lexicons/ invented employers, titles and departments, the denylist,
and the clinical and accusatory vocabularies this pack refuses
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, including both boundary checks
docs/ the reference dataset record, the verification record, and
what this pack does and does not claim under KVKK
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.
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.
Release v0.2.0 records the owner's sign-off for the baseline special rate, the full special-category template subset, publication of the reference datasets, and the declared dataset license.
Contributions under the Developer Certificate of Origin 1.1, no contributor license agreement. See CONTRIBUTING.md, GOVERNANCE.md, and SECURITY.md.
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 · banking · insurance
