{
  "schema_version": "1.0.0",
  "taxonomy_version": "1.0.0",
  "sources": {
    "sample-training-data": {
      "enabled": true,
      "format": "canonical_jsonl",
      "license_status": "author_provided",
      "notes": "Already uses the canonical schema and is still fully revalidated during every build."
    },
    "openpii-1.5m": {
      "enabled": true,
      "format": "openpii_jsonl",
      "revision": "a785eb528e28be2693c3718a27e066970de5dadb",
      "license_status": "CC-BY-4.0",
      "test_holdout_permille": 5,
      "benchmark_tier": "held_out_same_generator",
      "label_map": {
        "ACCOUNTNUM": {"type": "BANK_ACCOUNT_NUMBER", "subtype": null},
        "AGE": {"type": "AGE", "subtype": null},
        "AIRPORTCODE": null,
        "ALLERGIES": null,
        "AMOUNT": null,
        "BANKNAME": null,
        "BUILDINGNUM": {"type": "STREET_ADDRESS", "subtype": "building_number"},
        "CITY": {"type": "CITY", "subtype": "city"},
        "COUNTRY": {"type": "COUNTRY", "subtype": null},
        "CREDITCARDNUMBER": {"type": "PAYMENT_CARD_NUMBER", "subtype": null},
        "CURRENCY": null,
        "DATE": {"type": "DATE_TIME", "subtype": "unspecified_date"},
        "DOCTORNAME": {"type": "PERSON_NAME", "subtype": "clinician"},
        "DRIVERLICENSENUM": {"type": "DRIVER_LICENSE_NUMBER", "subtype": null},
        "EMAIL": {"type": "EMAIL_ADDRESS", "subtype": null},
        "GENDER": null,
        "GIVENNAME": {"type": "PERSON_NAME", "subtype": "first"},
        "HEIGHT": null,
        "HOSPITALNAME": null,
        "IDCARDNUM": {"type": "GOVERNMENT_ID", "subtype": "identity_card"},
        "IPV4": {"type": "IP_ADDRESS", "subtype": "ipv4"},
        "JOBTITLE": null,
        "JOBTYPE": null,
        "MACADDRESS": {"type": "MAC_ADDRESS", "subtype": null},
        "MEDICATION": null,
        "ORGANISATION": null,
        "PASSPORTNUM": {"type": "PASSPORT_NUMBER", "subtype": null},
        "PASSWORD": null,
        "POLICYNUM": {"type": "OTHER_UNIQUE_ID", "subtype": "policy_number"},
        "SALARY": null,
        "SEX": null,
        "SOCIALNUM": {"type": "GOVERNMENT_ID", "subtype": "social_number"},
        "STREET": {"type": "STREET_ADDRESS", "subtype": "street"},
        "SURNAME": {"type": "PERSON_NAME", "subtype": "last"},
        "TAXNUM": {"type": "TAX_ID", "subtype": null},
        "TELEPHONENUM": {"type": "PHONE_NUMBER", "subtype": null},
        "TIME": {"type": "DATE_TIME", "subtype": "time"},
        "TIMEZONE": null,
        "TITLE": null,
        "URL": {"type": "URL", "subtype": null},
        "USERNAME": {"type": "USERNAME", "subtype": null},
        "WEIGHT": null,
        "ZIPCODE": {"type": "POSTAL_CODE", "subtype": null}
      },
      "notes": "Every label observed in the pinned revision has an explicit decision. DATE remains generic DATE_TIME; it is never guessed to be DATE_OF_BIRTH. Sensitive attributes that are not identifiers in taxonomy v1 are explicitly excluded. SOCIALNUM maps to the generic GOVERNMENT_ID because only 3% of its occurrences are US-region; assigning the US-specific SOCIAL_SECURITY_NUMBER is left to the rule layer, which validates SSN structure and context."
    },
    "tab": {
      "enabled": true,
      "format": "tab_standoff_json",
      "revision": "558e09e26d6b36f5f78440074e6a233946d98bd9",
      "license_status": "MIT",
      "files": ["echr_test.json"],
      "split_assignment": "test",
      "benchmark_tier": "external_human_annotated",
      "adjudication": "majority",
      "strip_leading_titles": true,
      "label_map": {
        "PERSON": {"type": "PERSON_NAME", "subtype": null},
        "DATETIME": {"type": "DATE_TIME", "subtype": null},
        "CODE": null,
        "LOC": null,
        "ORG": null,
        "DEM": null,
        "QUANTITY": null,
        "MISC": null
      },
      "notes": "Independent human-annotated benchmark: 127 quality-checked ECHR judgments. Documents carry 1-10 competing annotator sets, adjudicated by majority vote after boundary normalization. TAB guidelines include honorifics in PERSON spans (80.3% of mentions) whereas taxonomy v1 excludes titles, so leading honorifics are stripped before voting. Only PERSON and DATETIME are mapped, so rows carry evaluated_types and the evaluator must ignore predictions of other types on these documents."
    },
    "physionet-deid-resources": {
      "enabled": false,
      "format": "recognizer_resources",
      "license_status": "resources_only",
      "notes": "Contains dictionaries/software/annotation resources but not the controlled source note text required for model rows."
    },
    "wikiann": {
      "enabled": false,
      "format": "token_tags_parquet",
      "license_status": "blocked_pending_license_review",
      "notes": "Quarantined. It is general silver NER, not PII gold data."
    },
    "multinerd": {
      "enabled": false,
      "format": "token_tags_jsonl",
      "license_status": "noncommercial_research_only",
      "notes": "Quarantined. It cannot enter the default commercial-capable training build."
    }
  }
}