sosa-pii-ner-th-v1.1.0

Thai regional PII span detection model. Fine-tune of urchade/gliner_large-v2.1 on synthetic Thai-language PII data. Deployed in SOSA DevOps as the TH-region sidecar NER uplift layer.

Labels

Label	Format	Description
th_national_id	13 digits	Thai national ID (บัตรประชาชน), DOPA-issued
th_tax_id	13 digits	Thai tax ID (เลขประจำตัวผู้เสียภาษี), Revenue Department
th_phone_local	+66 or 0X format	Thai mobile and landline numbers
th_health_hrn	HN-XXXXXX or numeric	Thai hospital record numbers (trained, not yet in eval)

Global labels also detected: email, phone_e164, credit_card, passport_generic, ipv4_public

Eval Results (v0.5.0 checkpoint — scoped TH+global, 606 examples)

Label	P	R	F1	Gate
th_national_id	0.9561	0.9083	0.9316	PASS (≥0.93)
th_phone_local	0.9195	0.8649	0.8914	PASS (≥0.85)
th_tax_id	0.7914	1.0000	0.8835	PASS (≥0.80)
th_health_hrn	—	—	—	not in test set

Training

• Base model: urchade/gliner_large-v2.1 (Apache 2.0)
• Hardware: RunPod A40 48GB, EU-SE-1
• Steps: 10,000 | Batch: 8 | Wall time: ~47 min
• Dataset: D-TH-3-v2 + D-TH-4 merged — 3,797 effective positives (negatives stripped per gliner_large-v2.1 collator requirement, re-encoded as sentinel spans)
• Internal checkpoint: sosa-pii-ner-th-v0.5.0
• Weights SHA-256: 699fcd4fc3981e1db47f840aa8c2b59be8ffbb0b2afe607f964e74d6a010474f

Training history (runs to reach v1.1.0)

Run	Dataset	NID F1	TAX F1	Status
v0.2.0 (v1.0.0)	D-TH-1	0.9735	0.9776	PASS — shipped as v1.0.0
v0.3.0	+D-TH-3	0.9342	0.7482	BLOCKED — TAX FP from comms-anchor negatives
v0.4.0	D-TH-3-v2	0.9336	0.7899	BLOCKED — TAX FP 57
v0.5.0	+D-TH-4	0.9316	0.8835	PASS — v1.1.0

Root cause of regression (resolved): 11 values appeared as both th_national_id and th_tax_id positives in training data (L7 value contamination). D-TH-4 AP-5 disambiguation pairs resolved the collision. TAX FP dropped from 64 → 29.

Known Limitations

• th_health_hrn trained (52 examples) but not in test set — F1 unmeasured
• th_tax_id precision 0.7914 — TAX FP still present on ambiguous 13-digit values without strong context signal (handled partially by context-domain training)
• Post-inference: no checksum filter on TH NID (format validation in sidecar Rust layer)

Integration (SOSA DevOps)

Loaded by sidecar when active_layers includes country:TH. Spans merged into Rust scrubber output via merge_sidecar_spans. Not a replacement for the always-on Rust regex layer — contextual uplift only.

License

Apache 2.0. Fine-tune artifact — attribution to urchade/gliner_large-v2.1 required.

Release History

Version	Date	Notes
v1.0.0	2026-05-29	Initial regional release
v1.1.0	2026-05-30	TAX precision recovery (+0.0835 F1), NID/TAX disambiguation pairs