temsa/OpenMed-PPSN-v5

OpenMed PPSN v5

Token classification model derived from OpenMed/OpenMed-PII-SuperClinical-Large-434M-v1 with added B-PPSN / I-PPSN support for Irish PPS Number detection.

What this release contains

• Full model weights (model.safetensors) with original OpenMed labels + PPSN labels.
• label_meta.json with label mapping and provenance.
• Repro/eval artifacts:
  • irish_ppsn_regression_v5.jsonl
  • eval_manual_irish_v5_raw.json
  • eval_hybrid_v5_no_plausible.json
  • ab_non_ppsn_v5_baseline.json

Key results

On irish_ppsn_regression_v5:

• Raw model-only PPSN performance: P=0.7778 R=0.8750 F1=0.8235.
• Recommended strict hybrid mode (model + checksum + span normalization, no plausible fallback):
  • P=1.0000 R=1.0000 F1=1.0000.

Non-PPSN retention check vs base OpenMed (A/B):

• Real-text behavior agreement F1: 1.0000.
• Entity density unchanged in the sampled real-text proxy set.

Recommended production usage

For best PPSN reliability, run this model with a strict hybrid post-processor:

1. Keep model detections for all labels.
2. For PPSN specifically, normalize/expand spans to full PPSN candidates.
3. Validate PPSN with checksum.
4. Disable broad "plausible" fallback in production (no-plausible-ppsn).

This repository includes only the model artifacts; hybrid scripts are in the companion codebase used to train/evaluate this release.

Limitations

• The Irish regression set is small and targeted; additional domain-specific validation is required before regulated deployment.
• PPSN detection in noisy OCR/ASR or heavily malformed text may require extra hardening.

License and attribution

• This derivative release is distributed under Apache-2.0, consistent with the base model license tag.
• Base model: OpenMed/OpenMed-PII-SuperClinical-Large-434M-v1.
• Training/evaluation used synthetic and augmented data sources, including nvidia/Nemotron-PII and joelniklaus/mapa.
• See NOTICE for attribution details.

QA quick check

• Load this model as a standard AutoModelForTokenClassification checkpoint.
• Run against irish_ppsn_regression_v5.jsonl.
• Confirm results in eval_manual_irish_v5_raw.json (raw) and eval_hybrid_v5_no_plausible.json (strict hybrid).