sosa-pii-ner-no-v1.0.0

Regional PII named entity recognition model for Norwegian 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 (459M parameters, DeBERTa-v3-large encoder) on synthetic Norwegian PII data. Detects fødselsnummer, kontonummer, organisasjonsnummer, and local phone numbers in natural Norwegian and English workplace text. Designed to complement a context-gated regex layer — not replace it.

Intended use: Local PII detection within the SOSA DevOps Privacy Filter sidecar. Text never leaves the user's machine. Not intended for cloud inference or batch processing of real patient/citizen data.

---

Labels

Label	Description	Format	Validator
no_national_id	Norwegian fødselsnummer / personnummer	11 digits DDMMYYNNNCC	Modulus-11 checksum
no_bank_account	Norwegian kontonummer	11 digits XXXX.XX.XXXXX	MOD-11 checksum
no_org_number	Norwegian organisasjonsnummer	9 digits	MOD-11 checksum
no_phone_local	Norwegian local phone (+47 prefix)	8 digits	Prefix range validation

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

---

Evaluation — v1.0.0 gate results

Methodology: Dual-basis evaluation. no_national_id gated on test_no.jsonl (150 gold NIDs all passing Modulus-11, 150 negatives — purpose-built to exclude synthetic non-checksummed values present in the shared multi-region test set). no_bank_account gated on the shared test set scoped to NO examples (270 examples). Post-inference Modulus-11 filter active in both evals.

Label	Test set	P	R	F1	Gate
no_national_id	test_no.jsonl (150 gold, checksummed)	1.000	0.807	0.893	≥ 0.85 ✅
no_bank_account	test.jsonl NO-scoped (270 ex)	0.818	0.982	0.893	≥ 0.80 ✅

Training: 12 runs. Best state T7 — D-NO-6 dataset (2,842 positives, 45% neg ratio), 10,000 steps, A40 GPU, 2026-05-29.

---

Production deployment notes

⚠️ Modulus-11 checksum validator must be co-deployed. Without it, high-confidence FP on 11-digit reference numbers reach output. Reference implementation: src-tauri/sidecar/validate_no_nid_checksum.py in the SOSA DevOps source repo.

---

Limitations

• Recall gap on structured/CSV contexts (R=0.807): NIDs in "field","value" format are underdetected. A format-regex + checksum layer covers this gap in production. Do not add structured-format training data — regression confirmed across multiple runs (T12).
• NID/bank 11-digit collision: Model fires no_bank_account on NID-context spans. A bank MOD-11 post-filter mitigates this in production.
• Context-gated: Bare values without surrounding context are unreliable. Use as recall-uplift on top of the regex layer, not standalone.
• Language: Norwegian Bokmål / Norwegian-English mixed documents.

---

Training data

Synthetic Norwegian PII examples only. No real citizen, patient, or employee data was used at any stage of training or evaluation. Negatives include realistic Norwegian workplace reference numbers (ansattnummer, ordrenummer, kundenummer) and multilingual adversarial strings.

---

Integrity

pytorch_model.bin SHA-256:

6bcde91df5b9301810a5e05988fbab91592f8570e5ebe125219c3bb4957f1836

---

License

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