temsa
/

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

Token Classification

de-identification

Model card Files Files and versions

OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc7 / eval /benchmark_summary.md

temsa's picture

Publish rc7 with spec-driven scanner release

32bcb86 verified 6 months ago

|

History Blame Contribute Delete

1.03 kB

Benchmark Summary

What Changed

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

The change is implementation transparency and maintainability:

scanner candidate extraction is defined in scanner_spec.yaml
runtime scanner data is generated into irish_core_generated_scanner_spec.py
semantic validation remains explicit and deterministic in code
release Python files no longer depend on regex
no regex-based candidate extraction is used in the repair layer

Full Checkpoint

Suite	rc6	rc7
Irish core manual	1.0000	1.0000
Phone / passport / finance	1.0000	1.0000
Finance boundary repair	1.0000	1.0000
Gaelic weak-context PPSN	1.0000	1.0000

ONNX q8

Suite	rc6	rc7
Irish core manual	0.9934	0.9934
Phone / passport / finance	1.0000	1.0000
Finance boundary repair	1.0000	1.0000
Gaelic weak-context PPSN	1.0000	1.0000