IndiaPII-Bench v1.0

A synthetic, adversarial, labelled benchmark for Indian PII detection: Aadhaar, PAN, GSTIN, IFSC, UPI VPA, ABHA, Indian mobile numbers, addresses, names, and more. Built for MaskFlow, the free, open-source PII masking library for LLM pipelines, but usable to evaluate any Indian-PII detector.

⚠️ Synthetic data — read before using

Every identifier in this dataset is synthetic. Aadhaar and GSTIN values are generated with mathematically valid checksums (Verhoeff, GSTIN mod-36); PANs, IFSC codes, and the rest are structurally valid. None of them belong to any real person, business, or account. Values are generated at random and cross-checked against maskflow-pack-india's own validators as part of the build (self_check() in generate.py) — proof that "checksum-valid" is actually true of what was generated, not merely asserted.

Dataset summary

• 2,000 documents, 13,468 labelled entity spans
• 7 document domains: support ticket, KYC form, insurance claim, medical note, HR record, bank chat transcript, loan application
• Deterministic build (seed 20260827) — anyone can regenerate this exact corpus
• Hinglish / code-mixed text, Devanagari + Latin script mixing, and messaging-style noise in the conversational domains (bank chat, support ticket)
• Labelled hard negatives: PII-shaped values that are not real PII (a non-Verhoeff 12-digit number, a PAN-shaped invoice number, a VPA-shaped email, order IDs, timestamps) — these are what make precision measurable, not just recall

Documents per domain

Domain	Count
bank_chat	285
hr_record	286
insurance_claim	286
kyc_form	286
loan_application	285
medical_note	286
support_ticket	286

Entity taxonomy

Positive labels (real PII, scored for recall)

Label	Description	Spans
AADHAAR	12-digit Indian national ID (Aadhaar), may be spaced/hyphenated in groups of 4	857
AADHAAR_MASKED	Aadhaar number with first 8 digits masked, only last 4 digits visible	285
ABHA_ADDRESS	Ayushman Bharat Health Account address, email-shaped (e.g. name@abdm)	141
ABHA_NUMBER	14-digit Ayushman Bharat Health Account number	286
BANK_ACCOUNT_IN	Indian bank account number, 9–18 digits	1,142
DRIVING_LICENCE	Indian driving licence number (state code + digits)	286
GSTIN	15-character Goods and Services Tax Identification Number	114
IFSC	11-character bank branch code (4 letters + 0 + 6 alphanumeric)	1,142
INDIAN_ADDRESS	A residential/postal address in India (street, locality, city, state)	1,230
INDIAN_MOBILE	Indian mobile phone number, 10 digits, optionally with +91 prefix	1,790
INDIAN_PASSPORT	Indian passport number (1 letter + 7 digits) or its MRZ block	101
PAN	10-character Permanent Account Number (5 letters + 4 digits + 1 letter)	857
PERSON_NAME	A person's full name	2,001
PIN_CODE	6-digit Indian postal PIN code	857
UPI_VPA	UPI Virtual Payment Address, looks like username@bank-handle	571
VEHICLE_REG	Indian vehicle registration number (state code + district + series + digits)	286
VOTER_ID	Voter ID / EPIC number, 3 letters + 7 digits	119

Hard-negative labels (PII-shaped, never real PII, scored against false positives)

Label	Description	Spans
NON_VERHOEFF_AADHAAR_SHAPED	A 12-digit number formatted like Aadhaar but failing the Verhoeff checksum	63
ORDER_ID_SHAPED	An order/reference number that resembles a structured identifier	564
PAN_SHAPED_INVOICE_NO	An invoice number in PAN's 5-letters-4-digits-1-letter shape	107
TIMESTAMP_SHAPED	An ISO-style timestamp long enough to be mistaken for a numeric ID	572
VPA_SHAPED_EMAIL	An email address in local@domain form, shaped like a UPI VPA	97

A detector that flags a hard-negative span is not penalized for missing it (it was never gold), but is penalized for the false positive it produced — this is what makes the benchmark's precision numbers meaningful, not just its recall.

Data format

One JSON object per line:

{
  "id": "indiapii-v1.0-00001",
  "text": "...",
  "domain": "insurance_claim",
  "lang": "en",
  "entities": [
    {"start": 47, "end": 64, "label": "PERSON_NAME", "value_class": "positive"}
  ]
}

entities[].value_class is "positive" (real, gold PII — scored for recall) or "hard_negative" (PII-shaped decoy — scored only against false positives, never gold).

Generation method

Built by a deterministic template-and-identifier generator (bench/indiapii/generator/): domain templates are filled with synthetic identifiers generated to pass (positives) or deliberately fail (hard negatives) each type's real validator, then noise (Hinglish code-mixing, script mixing, messaging-style abbreviation) is layered on top. A single random.Random(seed) instance threads through the whole run, so --seed 20260827 --count 2000 reproduces this exact corpus byte-for-byte.

Known limitations

• Realism noise is uneven across domains. The conversational domains (bank chat, support ticket) carry real Hinglish code-mixing and typos; the structured-form domains (KYC form, loan application, insurance claim) are still clean Field: value layouts with little noise applied. Treat scores on the form domains as an upper bound relative to real-world messiness.
• Synthetic text, generated from templates — even with noise applied, it will not fully capture the diversity of real documents.
• English and Hinglish only; no dedicated Devanagari-only documents.

Intended use

Evaluating and comparing Indian-PII detectors (regex/checksum, NER, LLM-based) — this is the dataset behind MaskFlow's own published benchmark (bench/reports/indiapii-v1.0/results.md). Not intended, and not suitable, for training a model to generate real-looking Indian identifiers.

Reproduction

git clone https://github.com/maskflow/maskflow
cd maskflow
uv sync --all-extras
uv run python bench/indiapii/generator/generate.py --count 2000 \
    --out bench/indiapii/data/indiapii-v1.0.jsonl

Score any detector against it (MaskFlow's own detector, no extra setup needed):

pip install 'maskflow-cli[bench]'
maskflow bench --my-data bench/indiapii/data/indiapii-v1.0.jsonl

Citation

@dataset{maskflow_indiapii_bench_2026,
  title        = {IndiaPII-Bench v1.0},
  author       = {Sodani, Somya},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/maskflow-ai/indiapii-bench}},
  license      = {CC-BY-4.0}
}

License

CC-BY-4.0. Attribution: MaskFlow project (https://github.com/maskflow/maskflow).