cagrigungor/turkish-pii-masking-benchmark

Turkish PII Masking Benchmark (1,000 test cases)

A hand-built, fully synthetic benchmark for evaluating instruction-conditional PII masking in Turkish. Designed to stress-test small LLMs (0.3B-1B) fine-tuned for privacy filtering in banking/ERP user messages. All values are synthetic (no real persons); all sentence patterns were written specifically for this benchmark (no training-set overlap).

Task

Given instruction (the masking policy) and input, the model must produce expected_output:

kategori	n	policy
tam	350	mask ALL PII fields
beyaz_liste	200	mask ONLY the listed fields
kara_liste	150	mask everything EXCEPT the listed fields
kapsam_disi	100	requested tag absent from text -> output must equal input
negatif	200	no PII at all (trap sentences) -> output must equal input

Hard slices (ozellikler column)

kayit multi-field records (5-9 tags) · caps ALL-UPPERCASE · uzun long inputs · ekli Turkish suffix-attached PII ("Cem Aslan'ın", vowel-harmony) · sozle spelled-out numbers ("yedi bir dokuz yedi") · cok_kisi multiple persons · tuzak PII-lookalike negatives (invoice numbers, city-named projects, PII words without PII values) · olmayan_etiket_talebi instruction mentions a tag not in the text

Columns

id, kategori, ozellikler, instruction, input, expected_output, metindeki_etiketler, maskelenecek_etiketler

Covers all 53 mask tags ([TCKN], [AD], [IBAN], [TEL], [SAGLIK], ... [YAS], [CINSIYET], [UYRUK]).

Also on Kaggle: https://www.kaggle.com/datasets/aiwithcagri/turkish-pii-masking-benchmark

Scoring

import pandas as pd
df = pd.read_csv("benchmark_1000.csv")
df["dogru"] = [model(r.instruction, r.input) == r.expected_output for r in df.itertuples()]
print(df.groupby("kategori").dogru.mean())

Exact match is the intended metric; per-slice breakdown via kategori x ozellikler.