training_sources.json

{
  "base_model": "OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1",
  "current_public_reference": "temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v1",
  "previous_rc_reference": "temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc1",
  "release_purpose": "Second v2 release candidate for the IrishCorePII line, focused on exact QA regressions around passport numbers, bank routing numbers, and Irish-format phone handling while preserving weak-context PPSN performance.",
  "recommended_thresholds": {
    "ppsn_min_score": 0.5,
    "other_min_score": 0.35
  },
  "training_mix_summary": [
    {
      "component": "exact QA-style passport, sort code, and compact-phone hardening mix",
      "weight": 3.0
    },
    {
      "component": "earlier Irish numeric QA hardening mix",
      "weight": 2.0
    },
    {
      "component": "broader Irish gap hardening replay",
      "weight": 2.0
    },
    {
      "component": "weak-context PPSN replay mix",
      "weight": 3.0
    },
    {
      "component": "broader Irish core PII replay mix",
      "weight": 2.0
    }
  ],
  "upstream_attribution": [
    {
      "name": "joelniklaus/mapa",
      "license": "CC-BY-4.0"
    },
    {
      "name": "gretelai/synthetic_pii_finance_multilingual",
      "license": "Apache-2.0"
    }
  ],
  "known_limitations": [
    "Standalone compact Irish mobile numbers remain weaker than spaced forms.",
    "Passport values written as prefix plus space plus digits can still lose the two-letter prefix span."
  ]
}