sosa-pii-ner-sg-v1.0.0

Regional PII named entity recognition model for Singapore documents. Part of the SOSA DevOps Privacy Filter — a local-first, privacy-preserving AI runtime for developers. Weights are Apache 2.0. No cloud required.

🔗 Source code: SOSA DevOps on GitHub 🌐 Product: sovereignsystems.cc

---

Model summary

Fine-tuned from urchade/gliner_large-v2.1 on synthetic Singapore PII data. Detects NRIC/FIN numbers, Unique Entity Numbers (UEN), Singapore local phone numbers, and healthcare reference numbers (HRN) in English documents.

Intended use: Local PII detection within the SOSA DevOps Privacy Filter sidecar. Text never leaves the user's machine.

---

Labels

Label	Description	Format	Validator
sg_nric_fin	Singapore NRIC / FIN	Letter + 7 digits + check letter (S/T/F/G/M)	Prefix + check character
sg_uen	Unique Entity Number	9–10 chars (numeric + alpha suffix, or T-prefix)	Format validation
sg_phone_local	Singapore local phone	8 digits, starting 3/6/8/9	Prefix validation
sg_health_hrn	Healthcare reference number	Alphanumeric 10–21 chars	Format validation

Global labels also carried (defence-in-depth): email, phone_e164, credit_card, passport_generic, ipv4_public

---

Evaluation — v1.0.0 gate results

Label	F1	Gate
sg_nric_fin	0.9524	≥ 0.85 ✅
sg_uen	0.7143	≥ 0.70 ✅

First-run gate pass (T1). Training: D-SG-1 dataset (1,157 positive examples), 10,000 steps, A40 GPU, 2026-05-28/29.

---

Limitations

• NRIC/UEN alphanumeric collision: Both labels use alphanumeric formats. Context (NRIC/FIN identity keywords vs UEN/company registration keywords) disambiguates.
• UEN precision: F1=0.7143 reflects limited UEN training examples. Adequate for production defence-in-depth; higher recall on NRIC is the primary detection surface.
• Context-gated: Bare values without surrounding context are unreliable.
• Language: English. Singapore-specific context templates (healthcare, HR, business registration).

---

Training data

Synthetic Singapore PII examples only. No real resident or patient data used.

---

Integrity

pytorch_model.bin SHA-256:

070db3ae5b22c7f40cf8037a604262f58a404d17bbf79dd69a26fe9f98edd91f

---

License

Apache 2.0 — inherited from urchade/gliner_large-v2.1. Fine-tuned by Sovereign Systems. See LICENSE.