IrishCore-GlobalPointer-ContextPII-135M-v1-rc29

IrishCore-GlobalPointer-ContextPII-135M-v1-rc29 is the current raw-only contextual PII masking release for Irish public-sector, HSE, and citizen-support flows.

It keeps the same weights and ONNX graph as the public temsa/IrishCore-GlobalPointer-ContextPII-135M-v1-rc25, but adds a role-aware decode policy for assistant/public prose and fixes the shared decoder inventory handling.

Positioning

• raw-only
• scanner-free
• validator-free
• single-pass GlobalPointer span extraction
• deterministic [PII:LABEL] replacement

What Changed

• adds a role-aware decode policy: assistant_public for assistant/public prose, default for user turns and search queries

• fixes the shared decoder so it respects the model label inventory instead of inventing contextual labels on non-context releases

• keeps the same weights and ONNX graph as the previous public checkpoint; this is a decode/runtime packaging release

• recovers Gaelic Árasán/Aonad + Teach/building-name + street address spans more completely

• restores full prefixed Irish city forms such as nDún Dealgan instead of suffix fragments

• keeps mixed contact phones after names/emails from being missed in assistant or user turns

• adds the fresh unseen contextual exact holdout v20 to the public release bundle

Benchmarks

ONNX q8

Evaluation Notes

• eval/q8_irish_gov_assistant_public_policy_v1.json: role-aware assistant/public policy suite; measures the conservative assistant profile on English and Irish assistant text
• eval/q8_irish_gov_user_contextual_policy_v1.json: role-aware user-contextual policy suite; measures the default profile on English and Irish user turns
• eval/q8_globalpointer_context_holdout_v18_exact.json: fresh unseen contextual holdout for public-office addresses, mixed public/private contacts, and Irish month-name negatives
• eval/q8_globalpointer_context_holdout_v19_exact.json: fresh unseen contextual holdout for standalone unit/building address spans and multilingual admin cue negatives
• eval/q8_globalpointer_context_holdout_v20_exact.json: fresh unseen contextual holdout for Gaelic Teach/building spans, prefixed Irish city forms, mixed contact phones, and admin-code negatives
• eval/q8_irish_gov_chatbot_gap_holdout_v2.json: Irish gov-chatbot holdout, 1.0000
• eval/q8_irish_gov_chatbot_redteam_v2.json: Irish gov-chatbot red-team suite, 0.9861

Role-Aware Usage

Use the model per message, not on a flattened whole transcript.

• --role user or --role search_query: uses the default contextual profile. This is the intended path for citizen turns and search queries.
• --role assistant: switches to the conservative assistant_public profile. This keeps direct structured identifiers such as PHONE_NUMBER, EMAIL, and POSTCODE, but suppresses contextual/demographic labels such as AGE, DATE_OF_BIRTH, CITY, COUNTY, STREET_ADDRESS, FIRST_NAME, and LAST_NAME in public explanatory prose.
• --profile ... overrides the role-derived default if you need explicit control.

Examples:

python3 inference_mask_onnx.py   --model temsa/IrishCore-GlobalPointer-ContextPII-135M-v1-rc29   --role user   --text "I'm 76 and I live at Rose Cottage, Galway Road, Loughrea, Galway County (H62 HF96)."

python3 inference_mask_onnx.py   --model temsa/IrishCore-GlobalPointer-ContextPII-135M-v1-rc29   --role assistant   --text "If you need to contact the Carer’s Allowance Section directly: Phone: 0818 927 999. Email: carers@example.gov.ie. Address: ... N39 E4E0."

For multi-turn transcripts, prefer the bundled role-aware helper instead of flattening the conversation into one string:

python3 inference_mask_messages_onnx.py \
  --model temsa/IrishCore-GlobalPointer-ContextPII-135M-v1-rc29 \
  --input-file messages.json \
  --json

python3 inference_mask_onnx.py --model temsa/IrishCore-GlobalPointer-ContextPII-135M-v1-rc29 --text "My PPSN is 1234567T and my address is Unit 6, Harbour Point, 19 Railway Road, Portlaoise, County Laois, R32 AB12."

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.