temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc8

OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc8

Raw-only Irish core PII release derived from OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1.

This repo does not require the scanner / validator layer used by temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc7. Both the full checkpoint and the ONNX q8 artifact use the same learned score-only decoder over token-presence and typed boundary heads.

Coverage

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

Included Variants

• —Full transformers checkpoint in the repo root
• —Unquantized ONNX export in onnx/model.onnx
• —Dynamic q8 ONNX artifact in onnx/model_quantized.onnx
• —inference_mask.py for the full checkpoint
• —inference_mask_onnx.py for the ONNX q8 artifact
• —common.py, model.py, and multitask_model.py implementing the raw-only decoder
• —benchmark files in eval/

Artifact sizes:

• —Full checkpoint: 515 MB (model.safetensors)
• —Dynamic q8 ONNX: 393 MB (onnx/model_quantized.onnx)

What Changed From rc7

rc7 achieves its public quality with a bundled scanner / validator inference stack.

rc8 removes that layer completely:

• —no regex-based candidate extraction
• —no checksum validator dependency at inference time
• —no separate scanner spec or generated scanner code
• —same compact mLiteClinical encoder family, but with a raw-only multi-head decoder

The tradeoff is explicit:

• —rc7 is still stronger on the broad manual Irish core suite
• —rc8 is easier to embed, simpler to maintain, and its ONNX q8 path stays very close to the full checkpoint

Architecture

rc8 keeps the DistilBERT-size encoder class of the 135M mLiteClinical base and adds:

• —a token-presence head for each released label
• —a typed start-boundary head
• —a typed end-boundary head
• —a score-only decoder that uses model scores, token offsets, continuity priors, and minimum-length priors from config

There is no scanner or external validator in the release path.

Design references that informed this direction:

• —Split-NER: https://aclanthology.org/2023.acl-short.36/
• —SpanNER: https://aclanthology.org/2021.acl-long.558/
• —Boundary Smoothing for Named Entity Recognition: https://aclanthology.org/2022.acl-long.490/
• —TinyBERT: https://aclanthology.org/2020.findings-emnlp.372/

How To Use It

Full checkpoint:

uv run python inference_mask.py \
  --model temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc8 \
  --min-score 0.5 \
  --text "My PPSN is 1234567TW, my Eircode is D02 X285, and my phone is 087 123 4567." \
  --json

Dynamic q8 ONNX:

uv run python inference_mask_onnx.py \
  --model temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc8 \
  --min-score 0.5 \
  --text "Please provide your passport NN5123456 and call me on 0851234567." \
  --json

Benchmarks

Main comparison:

Irish core label breakdown:

Dynamic q8 Artifact

Artifact paths:

• —unquantized: onnx/model.onnx
• —preprocessed: onnx/model.preprocessed.onnx
• —quantized: onnx/model_quantized.onnx

Quantization recipe used here:

• —ONNX pre-processing before quantization
• —ONNX Runtime dynamic int8
• —qint8
• —per_channel=true
• —op_types=MatMul,Gemm,Attention

For CPU deployment, the ONNX q8 artifact is the recommended default.

Limits

• —rc8 is raw-only. It intentionally gives up the scanner/validator stack used by rc7, so its broad manual-suite ceiling is lower.
• —The current remaining local misses are a bare 8-digit account-number case and one Gaelic phone-number case in the manual core suite.
• —If you need the exact rc8 behavior, use the bundled inference scripts or import decode_token_presence_segments from common.py.

License And Attribution

• —Release license: Apache-2.0
• —Base model: OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1
• —See NOTICE and training_sources.json for attribution and training details.

<!-- 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 -->