redact: on-device multilingual PII redaction

Detects and redacts personal data (names, addresses, emails, phone numbers, cards, IBANs, national IDs and more) in text across all 24 official EU languages (Latin, Greek and Cyrillic scripts). A BIOES token classifier plus a portable, dependency-free deterministic layer for structured IDs. The deployable model is ~13.7 MB (int4 ONNX); with the tokenizer the total on-device footprint is ~16 MB.

"Call Anna Kovács at anna@example.hu, IBAN GB29NWBK60161331926819" → "Call [GIVEN_NAME] [SURNAME] at [EMAIL], IBAN [BANK_ACCOUNT]"

Try it

• Live demo: desert-ant-labs/redact-demo: paste text and watch PII get highlighted or masked, fully in your browser.
• iOS / macOS / tvOS / visionOS: redact-swift: the Swift SDK (Swift Package Manager) with a built-in demo app. It bundles the compiled Core ML model below.
• Node / browser (JavaScript / TypeScript): redact-js (npm i @desert-ant-labs/redact): fetches the model from this repo (transformers.js) and caches it.
• JVM / Android (Kotlin): redact-kotlin (JitPack): ONNX Runtime with the bundled int4 model.

import Redact

let redact = Redact()
let r = try await redact.redaction(of: "Email Anna Kovács at anna@example.hu.")
r.redactedText   // "Email [GIVEN_NAME_1] [SURNAME_1] at [EMAIL_1]."

Files

File	Format	Size	Contents
redact.onnx	ONNX (int4, opset 21)	~13.7 MB	4-bit-quantized model, batch=1, ready for on-device runtimes
redact.mlmodelc	Compiled Core ML (4-bit)	~11.6 MB	Palettized model, ready to load on Apple platforms (used by redact-swift)
redact.pt	PyTorch checkpoint	~90 MB	Full-precision weights + config (for retraining / other runtimes)
config.json	JSON	tiny	Transformer + label config
tokenizer.json, tokenizer_config.json	JSON	~2.3 MB	EU-trimmed (31,475-piece) SentencePiece tokenizer (XLM-R lineage)
labels.json	JSON	tiny	BIOES id2label / label2id
redact_meta.json	JSON	tiny	Public labels, deterministic-owner labels, recommended thresholds, base-model info

Taxonomy (20 public labels)

GIVEN_NAME, SURNAME, STREET_NAME, BUILDING_NUMBER, SECONDARY_ADDRESS, CITY, STATE, ZIP_CODE, EMAIL, PHONE, CREDIT_CARD, BANK_ACCOUNT, ROUTING_NUMBER, IP_ADDRESS, URL, GOVERNMENT_ID, PASSPORT, DRIVERS_LICENSE, TAX_ID, SSN.

The deterministic layer additionally emits IMEI (device identifier), a deterministic-only label outside the 20-label neural head.

Architecture

• Encoder: Multilingual-MiniLM (XLM-R lineage) truncated to 6 layers with an EU-script-trimmed vocab (~23 M params), fine-tuned for BIOES tagging.
• Deterministic layer: a pure-stdlib post-processor owns high-confidence structured labels (email, URL, IP/MAC, card, IBAN/BIC, VIN, SSN, routing, tax id, government id, passport, driving licence, IMEI) with real validation (Luhn, ISO-13616 IBAN, ISO-7064, per-country checksums) and reconciles them with the model's contextual predictions. EU structured coverage includes checksum-validated national IDs for all 24 EU countries, all 27 EU VAT numbers, IMEI, and per-country driving-licence numbers. The same layer is ported byte-for-byte to the JS and Swift runtimes (span-for-span parity).
• Recommended runtime: min_score = 0.6, max_length = 256, stride = 64.

Benchmark

On a fair, all-label, 24-language evaluation (external WikiANN + MultiNERD for names/places, plus a neutral format-valid structured-PII set): fair composite leak-safe recall 88.8, typed F1 85.4, strict F1 71.0, best among on-device models by a wide margin, with better labeling accuracy and precision than models 100×+ its size.

License

Desert Ant Labs Source-Available License. Free for most apps; a commercial license is required at scale. Full terms are at the link. Licensing: licensing@desertant.ai.