---
language:
- en
- zh
license: apache-2.0
tags:
- ner
- pii
- gliner
- singapore
- privacy
- sosa
base_model: urchade/gliner_large-v2.1
---

# sosa-pii-ner-sg-v1.1.0

Singapore regional PII span detection model. Fine-tune of
[urchade/gliner_large-v2.1](https://huggingface.co/urchade/gliner_large-v2.1)
on synthetic Singapore-context PII data. Deployed in SOSA DevOps as the SG-region
sidecar NER uplift layer.

## Labels

| Label | Format | Description |
|---|---|---|
| sg_nric_fin | [STFG]XXXXXXX[A-Z] | Singapore NRIC/FIN (national ID / foreigner ID) |
| sg_uen | XXXXXXXXX[A-Z] (9-10 chars) | Singapore Unique Entity Number (business registry) |
| sg_phone_local | +65 XXXX XXXX | Singapore phone numbers |
| sg_health_hrn | SGH-YYYY-NNNNNN or numeric | Singapore hospital record numbers |

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

## Eval Results (v0.2.0 checkpoint — scoped SG+global, 219 examples)

| Label | P | R | F1 | Gate |
|---|---|---|---|---|
| sg_nric_fin | 0.9688 | 1.0000 | 0.9841 | PASS (≥0.85) |
| sg_uen | 0.5789 | 1.0000 | 0.7333 | PASS (≥0.70) |
| sg_health_hrn | 0.5172 | 1.0000 | 0.6818 | no gate (P2) |
| sg_phone_local | 0.6486 | 1.0000 | 0.7869 | no gate |

## 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: ~43 min
- **Dataset:** D-SG-2 + D-SG-2b merged — 1,299 effective positives
- **Internal checkpoint:** sosa-pii-ner-sg-v0.2.0
- **Weights SHA-256:** `32022cf0faa626cd23d0aa5c2a46a975cca488d1feb26cd518ee0eb72520e418`

### Training history (runs to reach v1.1.0)

| Run | Dataset | NRIC F1 | UEN F1 | Status |
|---|---|---|---|---|
| v0.1.0 (v1.0.0) | D-SG-1 | 0.9524 | 0.7143 | PASS — shipped as v1.0.0 |
| v0.2.0 | +D-SG-2b | 0.9841 | 0.7333 | PASS — v1.1.0 |

**D-SG-2b fixes:** UEN+phone, UEN+HRN, NRIC+HRN multi-entity co-occurrence gaps
(stress test failures SG-P11, SG-P13).

## Known Limitations

- `sg_uen` precision 0.5789 — 8 FP present (postal codes, reference numbers of UEN shape)
- `sg_health_hrn` precision 0.5172 — 28 FP (alphanumeric codes confused with HRN format)
- `sg_phone_local` precision 0.6486 — 39 FP (foreign numbers, short sequences)
- Precision gaps on UEN and HRN are P1/P2 — targeted for D-SG-3 in a future slice

## Integration (SOSA DevOps)

Loaded by sidecar when `active_layers` includes `country:SG`. 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](https://huggingface.co/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 | Multi-entity co-occurrence fixes (UEN+phone, UEN+HRN, NRIC+HRN) |