temsa/IrishCore-GlobalPointer-135M-v1-rc4

IrishCore-GlobalPointer-135M-v1-rc4

IrishCore-GlobalPointer-135M-v1-rc4 is a raw-only Irish PII masking model derived from OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1.

It targets explicit masking with placeholders such as [PII:PPSN], [PII:EMAIL], and [PII:POSTCODE] for:

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

The main target is English plus Irish Gaelic text in citizen-support, public-sector, and HSE-style flows. The repo ships both the full transformers checkpoint and a dynamic q8 ONNX artifact for CPU deployment.

What This Model Is

This release is a compact span-matrix extractor built on top of the OpenMed mLiteClinical encoder.

Short version:

• —Base OpenMed: plain BIO token classification
• —DiffMask: token-span extraction with per-token presence plus boundary heads
• —GlobalPointer: direct typed span scoring over the token grid

This release is fully raw-only:

• —no external regex scanner layer
• —no external checksum validator layer
• —only a bundled lightweight span-repair stage inside common.py

The published behavior is fully defined by the weights plus the bundled decode/repair logic in common.py.

Architecture

• —Encoder: DistilBERT-size encoder from OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1
• —Span head: GlobalPointer-style typed span matrix
• —Positional encoding inside the span head: RoPE
• —Training objective: positive-weighted BCE with hard-negative mining over valid upper-triangular spans
• —Runtime: one forward pass plus span decoding, with lightweight context-aware repairs for a few boundary-sensitive labels

Unlike the DiffMask line, this release does not use diffusion-style training. It directly predicts whether each (start_token, end_token) pair is an entity span for each released label.

Why This Line Exists

The main reason for this architecture change was to improve the cases where earlier raw-only lines were fragile:

• —mixed messages containing 2-3 PII types together
• —Eircode and IBAN/card boundaries in realistic support text
• —PPSN and phone/entity overlap cases
• —broader generalization on unseen nearby phrasing

The tradeoff is straightforward:

• —accuracy is materially better on the hard Irish/UAT suites
• —CPU q8 speed is slower than the DiffMask q8 line

Included Artifacts

• —full transformers checkpoint in the repo root
• —dynamic q8 ONNX artifact in onnx/model_quantized.onnx
• —inference_mask.py
• —inference_mask_onnx.py
• —benchmark JSON and summaries in eval/
• —release provenance in training_sources.json

Benchmark Summary

Gov Chatbot Red-Team v2

This Release

Closest Public Comparisons

Practical reading:

• —This release is the strongest raw-only checkpoint in this repo family on the hard Irish/UAT/holdout suites.
• —The q8 CPU artifact keeps the same accuracy on the main suites used for release.
• —The multilingual benchmark's overall score is lower than the PPSN label-only score because that suite penalizes extra name detections as false positives. On the label that matters there, PPSN, this release is strong.

Usage

Full checkpoint

python3 inference_mask.py \
  --model temsa/IrishCore-GlobalPointer-135M-v1-rc4 \
  --text "My PPSN is 1234567T, my Eircode is D02 XY45, and my phone is 087 123 4567." \
  --json

ONNX q8 CPU artifact

python3 inference_mask_onnx.py \
  --model temsa/IrishCore-GlobalPointer-135M-v1-rc4 \
  --text "My PPSN is 1234567T, my Eircode is D02 XY45, and my phone is 087 123 4567." \
  --json

Both scripts emit:

• —extracted spans with offsets and labels
• —masked text using [PII:LABEL]

How It Differs From Other OpenMed-Derived Lines

Compared with the original OpenMed base model:

• —this release does not use a BIO token head
• —it scores typed spans directly
• —it is tuned only for the released Irish core PII labels

Compared with DiffMask:

• —no diffusion-style masked denoising during training
• —no token-presence or boundary heads
• —direct span-matrix scoring instead of reconstructing spans from token and boundary logits

Compared with the old hybrid releases:

• —no external scanner or validator layer is required for the published behavior
• —deployment stays simple: encoder + span head + deterministic replacement

Limitations

• —It is slower on CPU than the DiffMask q8 release.
• —The multilingual benchmark is not a clean multilingual NER benchmark; it is a PPSN-focused suite where extra name detections count against the overall score.
• —If you need the fastest CPU-first public line rather than the strongest raw-only accuracy, compare this release against temsa/IrishCore-DiffMask-135M-v1-rc6.

References

• —DistilBERT: https://arxiv.org/abs/1910.01108
• —Global Pointer: Novel Efficient Span-based Approach for Named Entity Recognition: https://arxiv.org/abs/2208.03054
• —RoFormer: Enhanced Transformer with Rotary Position Embedding: https://arxiv.org/abs/2104.09864

License And Attribution

• —Release license: Apache-2.0
• —Base model: OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1
• —Dataset attribution is listed in NOTICE

<!-- 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 -->