Werea-co/Werea-KVKK-Bench

Werea-KVKK-Bench v2

Synthetic, reproducible data and benchmark assets for evidence-first Turkish privacy operations.

Contents

• —pii_synthetic.jsonl: fictional Turkish PII spans;
• —workflow_sft.jsonl: structured workflow-assistant conversations;
• —cases.jsonl: frozen rules/deadline/PII evaluation cases;
• —kvkk_rules.json: human-approval workflow definitions;
• —kvkk_controls.json: evidence-based institutional control catalog;
• —pii_taxonomy.json: Turkish personal/special-category data taxonomy;
• —official_sources.json: official URLs and short navigational summaries;
• —scripts/: deterministic generators, evaluators and gated training code.

Every generated record has synthetic: true or equivalent provenance. Email addresses use example.invalid; companies and people are fictional. No customer record, raw court file or copied full legal corpus is included.

The machine-verified benchmark is not a substitute for a lawyer-reviewed hidden test set. Production claims require an independent Turkish privacy professional to review source correctness, workflow coverage, abstention behaviour and real-document performance.

Generation

python scripts/build_synthetic.py
PYTHONPATH=src python benchmark/evaluate_core.py

Seed: 20260825. Rule version: 2026-08-25.2.

Intended use

Research, reproducible evaluation, structured-output training and privacy-tool prototyping. It must not be used to claim legal compliance or to automate consequential decisions without authorised human review.

Version 2 expands the catalog to 51 workflows and the synthetic PII set to 2,400 rows. Passing this benchmark measures conformance to the published synthetic contract only; it is not proof of legal completeness or real-world accuracy.