temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc2

temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc2

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

This repository should be evaluated against the public release and the previous RC:

• —current public release: temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v1
• —previous RC: temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc1
• —this repository: temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc2

The purpose of this RC is specific: keep the v2-rc1 weak-context PPSN recovery and improve the QA regressions reported afterwards for:

• —passport numbers like PA1234567, NN5123456, and some Irish Gaelic passport phrasing
• —Irish sort codes like 90-00-17, 90 01 18, and 900118
• —broader Irish-core stability relative to the public v1 release

This repo now also includes a bundled dynamic int8 ONNX export for CPU inference in onnx/model_quantized.onnx.

Included Variants

Coverage

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

Recommended Inference

Full checkpoint:

python3 inference_mask.py \
  --model temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc2 \
  --ppsn-min-score 0.5 \
  --other-min-score 0.35 \
  --text "My sort code is 90-00-17 for AIB." \
  --json

Fast CPU path with the bundled ONNX q8 artifact:

python3 inference_mask_onnx.py \
  --model temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc2 \
  --ppsn-min-score 0.5 \
  --other-min-score 0.35 \
  --text "Please provide your passport: NN5123456." \
  --json

Quantized Artifact

The bundled quantized artifact is:

• —onnx/model_quantized.onnx

It was benchmarked against the full v2-rc2 checkpoint and the public baselines at the same release thresholds.

Short reading:

• —exact QA: ONNX q8 is below the full checkpoint on passport boundary cases, but it keeps the routing and phone gains
• —English/Irish core suite: ONNX q8 is slightly higher overall than the full checkpoint on this suite
• —multilingual PPSN suite: ONNX q8 is below the full checkpoint and far below the public v1 PPSN-focused release
• —CPU throughput: ONNX q8 is much faster on the Irish core and edge suites, but not on every workload

Comparison To Public v1, v2-rc1, And ONNX q8

Exact QA suites:

Broader CPU benchmarks:

CPU throughput:

Practical Reading Of The ONNX q8 Results

• —The ONNX q8 artifact keeps the RC's user PPSN regression score at 1.0000 and the edge-suite overall score at 0.9500.
• —On the exact QA suites, ONNX q8 is still strong on bank routing and phone, but the full checkpoint is better on passport boundary cases.
• —On the English/Irish core suite used for this release, ONNX q8 is slightly better overall than the full checkpoint at the same thresholds.
• —The main quality cost of ONNX q8 in this release line is multilingual PPSN precision: 0.7887 for ONNX q8 vs 0.8038 for the full checkpoint, and both are below the public v1 PPSN-focused release (0.9940).

Known Limits

This is still a raw-model-only release candidate. QA should still test these carefully:

• —standalone compact Irish mobile numbers such as 0851234567
• —compact Irish mobiles in short contexts such as Call me on 0851234567 tomorrow.
• —passport values written as prefix + space + digits, such as Passport PA 1234567 was used to board the flight.

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 data created in this workspace

See NOTICE for attribution 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 -->