OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc7

Token-classification release for Irish core PII in English and Irish Gaelic.

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
• benchmark files in eval/

Coverage

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

What Changed From rc6

rc7 keeps the same checkpoint weights and the same bundled ONNX q8 artifact as temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc6.

The change is the scanner implementation:

• candidate extraction is now defined in a human-readable spec: scanner_spec.yaml
• generated runtime data lives in irish_core_generated_scanner_spec.py
• regeneration is explicit: python generate_scanner_spec.py
• semantic validation remains procedural in ppsn.py, eircode.py, and irish_core_decoder.py
• release Python files no longer depend on the third-party regex package
• the scanner layer no longer uses regex-based candidate extraction on untrusted text
• PPSN candidate extraction no longer consumes a following word-initial letter after whitespace, so cases like 1234567T a ... stay bounded to 1234567T

This is not a pure PEG grammar because some labels need semantic checks after lexical scanning:

• PPSN: checksum and suffix plausibility
• ACCOUNT_NUMBER: Irish IBAN structure and mod-97
• CREDIT_DEBIT_CARD: Luhn and grouped-card plausibility
• POSTCODE: Eircode routing-key and unique-identifier validation

So the public contract is:

1. human-readable lexical scan spec in scanner_spec.yaml
2. generated scanner configuration in irish_core_generated_scanner_spec.py
3. deterministic semantic validators in code

How To Use It

If you want the published rc7 behavior, use the bundled inference stack. A plain transformers token-classification pipeline will not reproduce the release behavior on numeric/boundary-heavy labels.

Full checkpoint:

uv run python inference_mask.py \
  --model temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc7 \
  --ppsn-min-score 0.55 \
  --other-min-score 0.50 \
  --text "Please provide your passport: NN5123456 and call me on 0851234567." \
  --json

Dynamic q8 ONNX:

uv run python inference_mask_onnx.py \
  --model temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc7 \
  --onnx-file onnx/model_quantized.onnx \
  --ppsn-min-score 0.55 \
  --other-min-score 0.50 \
  --text "My IBAN is IE29 AIBK 9311 5212 345678 and my PPSN is 1234567T." \
  --json

Key Benchmarks

Benchmarked with the spec-driven scanner stack shipped in this repo.

Full Checkpoint

Suite	F1
Irish core manual	1.0000
Phone / passport / finance	1.0000
Finance boundary repair	1.0000
QA Gaelic weak-context PPSN	1.0000

ONNX q8

Suite	F1
Irish core manual	0.9934
Phone / passport / finance	1.0000
Finance boundary repair	1.0000
QA Gaelic weak-context PPSN	1.0000

Irish Core Label Breakdown

Label	Full	ONNX q8
PPSN	1.0000	1.0000
PHONE_NUMBER	1.0000	1.0000
POSTCODE	1.0000	0.8571
PASSPORT_NUMBER	1.0000	1.0000
ACCOUNT_NUMBER	1.0000	1.0000
BANK_ROUTING_NUMBER	1.0000	1.0000
EMAIL	1.0000	1.0000
FIRST_NAME	1.0000	1.0000
LAST_NAME	1.0000	1.0000

Dynamic q8 Artifact

Artifact paths:

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

Quantization recipe used in this repo:

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

Limits

• POSTCODE remains the main q8 gap on the manual Irish core suite.
• The release contract includes the bundled decoder; raw pipeline(...) output is not the release behavior.
• The scanner spec covers lexical extraction; semantic validators remain in code by design.

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 release details.