sosa-pii-ner-th-v1.0.0

Regional PII named entity recognition model for Thai 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 Thai PII data. Detects Thai national IDs (เลขบัตรประชาชน), tax IDs, local phone numbers, and healthcare reference numbers in Thai and Thai-English mixed 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
th_national_id	Thai national ID (เลขบัตรประชาชน)	13 digits, first digit 1–8	Thai check digit (weighted sum mod 11)
th_tax_id	Thai tax identification number	13 digits, first digit 0–8	Length + prefix
th_phone_local	Thai local phone number	9–10 digits, 06/08/09/02/etc. prefix	Prefix validation
th_health_hrn	Thai hospital reference number (HRN)	Alphanumeric, 8–24 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
th_national_id	0.9735	≥ 0.85 ✅
th_tax_id	0.9776	≥ 0.80 ✅
th_phone_local	0.9814	≥ 0.85 ✅

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

---

Limitations

• NID/TAX 13-digit collision: Both labels share 13-digit format. Context (DOPA/civil keywords vs Revenue Department/VAT keywords) disambiguates. Validated via AP-5 disambiguation pairs in training data.
• Context-gated: Bare values without surrounding context are unreliable.
• Language: Thai and Thai-English mixed text.

---

Training data

Synthetic Thai PII examples only. No real citizen or patient data used.

---

Integrity

pytorch_model.bin SHA-256:

ca6219615d8e6e6f4da39762add863334abc5374ac53eb39b5fa7093f33efbc9

---

License

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