temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc4

Fourth QA release candidate for Irish core PII detection with OpenMed mLiteClinical.

This repository should be evaluated against:

• current public candidate: temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc3
• stable public release: temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v1
• this repository: temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc4

This RC improves the current public candidate on the main release-gating suites without reopening the PPSN false positives that were addressed after v2-rc3.

Included Variants

Variant	Artifact	Backend	Recommended Thresholds	Intended Use
Full checkpoint	repo root	transformers	ppsn=0.71, other=0.50	highest-fidelity evaluation and deployment
Quantized checkpoint	onnx/model_quantized.onnx	ONNX Runtime dynamic int8	ppsn=0.71, other=0.60	CPU-oriented deployment

Coverage

• PPSN
• account_number
• bank_routing_number
• credit_debit_card
• PASSPORT_NUMBER
• postcode
• phone_number
• email
• first_name
• last_name
• swift_bic

The main focus is English and Irish Gaelic handling for Irish administrative, citizen-support, and HSE-style text.

Recommended Inference

Full checkpoint:

uv run python inference_mask.py \
  --model temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc4 \
  --ppsn-min-score 0.71 \
  --other-min-score 0.50 \
  --text "My PPSN is 1234567T and my sort code is 90-00-17." \
  --json

Fast CPU path with the bundled ONNX q8 artifact:

uv run python inference_mask_onnx.py \
  --model temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc4 \
  --ppsn-min-score 0.71 \
  --other-min-score 0.60 \
  --text "Please provide your passport: NN5123456." \
  --json

If you prefer plain python3, install the dependencies from pyproject.toml first.

What Improved Versus temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc3

Full checkpoint:

• core suite F1: 0.9806 -> 0.9870
• overlap suite F1: 0.9429 -> 1.0000
• strict remaining IoU=1.0 F1: 0.4444 -> 0.6000
• multilingual PPSN-only F1: 0.9333 -> 0.9545
• matches the current public candidate on edge, numeric, gap, and user PPSN regression F1

Bundled ONNX q8:

• core suite F1: 0.9677 -> 0.9804
• multilingual PPSN-only F1: 0.9333 -> 0.9600
• keeps the overlap and strict remaining gains from v2-rc3
• q8 remains weaker than the full checkpoint on the small edge suite: 0.9474 vs 1.0000

Benchmark Table

Variant	Core	Edge	Numeric	Gap	User PPSN	Overlap	Strict Remaining IoU=1.0	Multilingual PPSN
temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc3 full	0.9806	1.0000	0.9333	0.9167	1.0000	0.9429	0.4444	0.9333
temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc4 full	0.9870	1.0000	0.9333	0.9167	1.0000	1.0000	0.6000	0.9545
temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc3 ONNX q8	0.9677	1.0000	0.9333	0.9167	1.0000	1.0000	0.6667	0.9333
temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc4 ONNX q8	0.9804	0.9474	0.9333	0.9167	1.0000	1.0000	0.6667	0.9600

Quantized Artifact

The bundled quantized artifact is:

• onnx/model_quantized.onnx

For this release line, the promoted q8 recipe remains the standard ONNX Runtime dynamic int8 export with per-channel quantization over MatMul, Gemm, and Attention.

Two alternatives were reviewed and not promoted for this model family:

• QAT in this DistilBERT token-classification stack
• Mezzanine's recent Qwen-focused weight transforms

Known Limits

This is still a raw token-classification release candidate without hybrid rule logic. QA should still test these carefully:

• Passport PA 1234567 was used to board the flight.
• Usaideadh pas PA 1234567 chun dul ar bord an eitilt.
• Call me on 0851234567 tomorrow.

Included Files

• full transformers checkpoint in the repo root
• dynamic int8 ONNX artifact in onnx/model_quantized.onnx
• inference_mask.py
• inference_mask_onnx.py
• qa_config.json
• training_sources.json
• benchmark summaries in eval/

License And Attribution

• release license: Apache-2.0
• base model: OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1
• upstream attributed data: joelniklaus/mapa, gretelai/synthetic_pii_finance_multilingual
• synthetic Irish training and replay data created in this workspace

See NOTICE for attribution details.