IrishCore-GlobalPointer-ContextPII-135M-v1-rc9

IrishCore-GlobalPointer-ContextPII-135M-v1-rc9

IrishCore-GlobalPointer-ContextPII-135M-v1-rc9 is the current expanded-label raw-only PII masking release for Irish public-sector, HSE, and citizen-support flows.

It keeps the same DistilBERT-size GlobalPointer span extractor family as temsa/IrishCore-GlobalPointer-ContextPII-135M-v1-rc8, but ships a stronger bundled decoder for contextual age phrasing, Gaelic ordinal DOB recovery, parenthesized landlines, and date-internal city/name overlap cleanup.

Context labels served by this line:

• STREET_ADDRESS
• CITY
• COUNTY
• DATE_OF_BIRTH
• AGE

Core labels retained:

• PPSN
• POSTCODE
• PHONE_NUMBER
• EMAIL
• PASSPORT_NUMBER
• ACCOUNT_NUMBER
• BANK_ROUTING_NUMBER
• SWIFT_BIC
• CREDIT_DEBIT_CARD
• FIRST_NAME
• LAST_NAME

Positioning

rc9 is a decoder-hardening release over rc8.

• weights unchanged
• ONNX graph unchanged
• no external scanner or validator added
• deployment path still single-pass span extraction plus deterministic [PII:LABEL] replacement

What changed in rc9:

• the decoder now recovers contextual age spans from phrasing like aged 43, 43-year-old, and 43 y/o
• the decoder now repairs Gaelic ordinal DOB spans such as an 3ú Márta 1991
• the decoder now repairs parenthesized Irish landlines like (01) 240 8787
• the decoder now prevents date-internal city/name overlap drift such as Aibrean leaking bre
• the new globalpointer_context_redteam_v3 suite is 1.0000 on both full and q8 paths; the public rc8 bundle scores 0.8718 on the same suite

Use this release when you need broader masking for Irish gov / HSE / citizen-support text, including user turns and assistant answers that contain:

• personal address fragments
• city / county
• date of birth
• age
• official callback numbers or public-service mailbox emails that still need masking in assistant output

If you only need the narrower Irish-core structured label set and want maximum CPU throughput, temsa/IrishCore-GlobalPointer-135M-v1-rc4 remains the faster option.

Architecture

• base encoder: OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1
• extractor head: GlobalPointer-style typed span matrix
• runtime: single-pass span extraction
• output policy: deterministic bracket masking, for example [PII:PPSN]
• deployment target: ONNX Runtime CPU with dynamic q8 per-channel quantization

This is not a generative model and does not rewrite text. It predicts typed spans and the provided inference scripts replace them with [PII:LABEL] placeholders.

Decoder Policy

The serving path is still raw-only in the sense that there is no external scanner or validator service. The repo does include bundled decoder repairs for highly structured or contextual spans:

• PPSN normalization and overlap repair
• passport cue-based repair
• contextual date-of-birth repair, including ordinal Gaelic date forms
• contextual age recovery from aged, year-old, and y/o phrasing
• contextual phone recovery, including Fón: cues, +353 (0) formats, and parenthesized landlines
• full-email recovery
• Eircode recovery
• contextual street-address recovery, including cue-led suffixless address blocks and longer apartment / house-name spans
• explicit county phrase recovery with county-over-name overlap cleanup
• prefixed Irish city-form recovery

These repairs live inside common.py and are part of the published inference path.

Benchmarks

ONNX q8

Full checkpoint

Comparison

Tradeoff:

• this expanded-label line is materially better on contextual Irish masking tasks, including age phrasing, Gaelic ordinal DOBs, landline formatting, and prefixed Gaelic city repairs
• it is slower than the core-only GlobalPointer line on CPU because it carries a broader label inventory and more decoder work

Evaluation Notes

Additional q8 release checks shipped in this repo:

• eval/q8_irish_numeric_qafix_v2.json: numeric false-positive guardrail suite, 1.0000 F1
• eval/q8_globalpointer_context_redteam_v1.json: contextual hardening suite for apartment-prefix street addresses, explicit County/Contae forms, and public-office address blocks, 1.0000 F1
• eval/q8_globalpointer_context_redteam_v2.json: contextual regression suite for born DOB phrasing, +353 (0) phone formatting, longer apartment / house-name address spans, and hUaimh, 1.0000 F1
• eval/q8_globalpointer_context_redteam_v3.json: contextual regression suite for age phrasing, Gaelic ordinal DOBs, parenthesized landlines, and date-internal city/name overlap cleanup, 1.0000 F1 in rc9 versus 0.8718 in the public rc8 bundle
• eval/q8_globalpointer_location_coverage_v1.json: broader Irish city/county/address coverage suite, 1.0000 F1
• eval/q8_globalpointer_location_coverage_v2.json: prefixed-Gaelic city-form suite, 1.0000 F1
• eval/q8_globalpointer_location_coverage_v3.json: additional prefixed-Gaelic city-form suite, 1.0000 F1

Other notes:

• globalpointer_demographic_patch_v2_test is the corrected held-out benchmark. The earlier v1 demographic patch contained invalid synthetic Eircodes in some rows.
• irish_gov_contact_policy_v1 is the policy-aligned assistant-output benchmark for this release.
• globalpointer_context_redteam_v3 is the new contextual regression benchmark for age phrasing, ordinal DOB forms, and parenthesized landlines.
• The legacy irish_gov_chatbot_redteam_v2 negatives still assume some public assistant contact details should not be masked. That assumption does not match this release's target policy.

Usage

Full checkpoint:

python3 inference_mask.py   --model temsa/IrishCore-GlobalPointer-ContextPII-135M-v1-rc9   --text "The claimant was aged 43, born on an 3ú Márta 1991, and can be reached on (01) 240 8787."

ONNX q8:

python3 inference_mask_onnx.py   --model temsa/IrishCore-GlobalPointer-ContextPII-135M-v1-rc9   --text "Seoladh: Apartment 4B, 12 Main Street, Galway, H91 F4E2."

Expected masking style:

The claimant was aged [PII:AGE], born on [PII:DATE_OF_BIRTH], and can be reached on [PII:PHONE_NUMBER].

Files

• model.safetensors: full checkpoint
• onnx/model_quantized.onnx: recommended CPU artifact
• inference_mask.py: full-checkpoint inference
• inference_mask_onnx.py: ONNX q8 inference
• eval/benchmark_summary.json: machine-readable benchmark summary
• training_sources.json: data provenance

Limitations

• This release is Irish-first. Multilingual overall precision is still pulled down by extra name detections outside the primary Irish target domain.
• The decoder deliberately prefers recall on structured Irish identifiers and contextual Irish masking. If you need a stricter non-Irish name policy, test on your own corpora before promoting beyond rc.

<!-- portfolio-comparison:start -->

Portfolio Comparison

Updated: 2026-03-16.

Use this section for the fastest public comparison across the temsa PII masking portfolio.

• The first core table only includes public checkpoints that ship both comparable q8 accuracy and q8 CPU throughput.
• The first PPSN table only includes public artifacts that ship comparable PPSN accuracy and CPU throughput.
• Missing cells in the archive tables mean the older release did not ship that metric in its public bundle.
• DiffMask rows use the reconciled clean_single_pass harness that matches the deployed runtime.
• GlobalPointer rows use the public raw-only span-matrix release bundle and its packaged q8 ONNX artifact.
• The same content is shipped as PORTFOLIO_COMPARISON.md inside each public model repo.

Irish Core PII: Comparable Public Checkpoints

Irish Core PII: Other Public Checkpoints

Finance-boundary q8 F1 is 1.0000 for OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc6, OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc7, OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc8, and all public IrishCore-DiffMask releases from rc1 to rc6. OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc5 ships 0.8750 on that public q8 suite.

PPSN-Only: Comparable Public Artifacts

PPSN-Only: Historical Public Checkpoints

If you need the strongest current raw-only Irish core model, start with IrishCore-GlobalPointer-135M-v1-rc4. If you need the fastest CPU-first raw-only line, compare it against IrishCore-DiffMask-135M-v1-rc6. If you need a PPSN-only artifact, compare the canonical fp32, fp16, and q8 variants of OpenMed-mLiteClinical-IrishPPSN-135M-v1 directly in the table above.

<!-- portfolio-comparison:end -->