training_sources.json

{
  "base_model": "OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1",
  "previous_public_candidate": "temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc4",
  "stable_public_reference": "temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v1",
  "release_purpose": "Fifth v2 release candidate for the IrishCorePII line. rc5 keeps the rc4 checkpoint weights but changes the public inference defaults and bundled dynamic q8 artifact to recover Gaelic weak-context PPSN cases and improve CPU q8 throughput.",
  "recommended_thresholds": {
    "full_checkpoint": {
      "ppsn_min_score": 0.55,
      "other_min_score": 0.5
    },
    "onnx_q8": {
      "ppsn_min_score": 0.55,
      "other_min_score": 0.5
    }
  },
  "inference_stack_notes": [
    "PPSN extraction uses the word-aligned decoder.",
    "General Irish core PII extraction uses the label-aware repair decoder for both full and ONNX q8 inference.",
    "The ONNX q8 artifact is built from a preprocessed ONNX export and then dynamically quantized with qint8 per-channel quantization over MatMul,Gemm,Attention."
  ],
  "training_mix_summary": [
    {
      "component": "same fine-tuned full checkpoint weights as rc4",
      "weight": 1.0
    },
    {
      "component": "updated inference calibration for weak-context Gaelic PPSN recovery",
      "weight": 1.0
    },
    {
      "component": "preprocessed ONNX export before dynamic q8 quantization",
      "weight": 1.0
    }
  ],
  "upstream_attribution": [
    {
      "name": "temsa/OpenMed-Irish-PPSN-Eircode-Spec-v1",
      "license": "Apache-2.0"
    },
    {
      "name": "temsa/OpenMed-Irish-CorePII-TrainMix-v1",
      "license": "CC-BY-4.0 composite"
    },
    {
      "name": "joelniklaus/mapa",
      "license": "CC-BY-4.0"
    },
    {
      "name": "gretelai/synthetic_pii_finance_multilingual",
      "license": "Apache-2.0"
    }
  ],
  "quantization_notes": {
    "promoted_q8_recipe": "ONNX Runtime dynamic int8 qint8 per-channel quantization over MatMul,Gemm,Attention after ONNX pre-processing with symbolic shape inference disabled.",
    "rejected_q8_recipes": [
      "non-per-channel dynamic int8",
      "MatMul,Gemm-only dynamic int8 without ONNX pre-processing"
    ]
  },
  "known_limitations": [
    "The full checkpoint still outperforms q8 on the finance-boundary suite.",
    "The strict remaining-gap suite is still weaker on q8 than on the full checkpoint.",
    "Grouped credit/debit card boundary cases remain the main shared weakness."
  ]
}