| {
|
| "schema_version": "1.0.0",
|
| "taxonomy_version": "1.0.0",
|
| "evidence_build": "prepared-training-data full build, OpenPII revision a785eb528e28be2693c3718a27e066970de5dadb",
|
| "thresholds": {
|
| "model_trained_min_entities": 10000,
|
| "experimental_min_entities": 100
|
| },
|
| "tiers": {
|
| "model_trained": {
|
| "description": "Sufficient span evidence to train and evaluate a contextual head.",
|
| "types": {
|
| "PERSON_NAME": 1569634,
|
| "DATE_TIME": 1419959,
|
| "EMAIL_ADDRESS": 881037,
|
| "CITY": 824794,
|
| "STREET_ADDRESS": 774697,
|
| "GOVERNMENT_ID": 668691,
|
| "PHONE_NUMBER": 660835,
|
| "AGE": 588563,
|
| "POSTAL_CODE": 506110,
|
| "PAYMENT_CARD_NUMBER": 374410,
|
| "DRIVER_LICENSE_NUMBER": 328122,
|
| "TAX_ID": 308681,
|
| "PASSPORT_NUMBER": 239124
|
| }
|
| },
|
| "experimental": {
|
| "description": "Some evidence but below the training threshold. Served only when the caller opts into experimental types.",
|
| "types": {
|
| "URL": 262,
|
| "COUNTRY": 214,
|
| "USERNAME": 101
|
| }
|
| },
|
| "rule_only": {
|
| "description": "Deterministic format or checksum detection. Excluded from the model label set by design, not by lack of data.",
|
| "types": {
|
| "IBAN": 1,
|
| "SWIFT_BIC": 1,
|
| "CRYPTO_WALLET": 1,
|
| "MAC_ADDRESS": 2,
|
| "IP_ADDRESS": 20,
|
| "GEO_COORDINATES": 1,
|
| "VEHICLE_IDENTIFIER": 1,
|
| "DEVICE_IDENTIFIER": 1,
|
| "AADHAAR_NUMBER": 1,
|
| "FAX_NUMBER": 1,
|
| "SOCIAL_SECURITY_NUMBER": 1
|
| }
|
| },
|
| "blocked": {
|
| "description": "Context-dependent types that need annotated data before they can be claimed. Sources are gated or not yet generated.",
|
| "types": {
|
| "MEDICAL_RECORD_NUMBER": 1,
|
| "HEALTH_PLAN_ID": 1,
|
| "PROVIDER_ID": 1,
|
| "DATE_OF_BIRTH": 1,
|
| "ADDRESS": 1,
|
| "COUNTY_DISTRICT": 1,
|
| "STATE_PROVINCE": 1,
|
| "LICENSE_PLATE": 1,
|
| "VOTER_ID": 1,
|
| "BIOMETRIC_IDENTIFIER": 1,
|
| "BANK_ACCOUNT_NUMBER": 47,
|
| "OTHER_UNIQUE_ID": 3
|
| }
|
| }
|
| },
|
| "training_notes": [
|
| "The model label set is exactly the model_trained tier; experimental types may be added only behind an explicit flag.",
|
| "Entities whose type is not in the model label set must be loss-masked, never relabelled as O. Relabelling teaches the model to suppress spans that rule recognizers are expected to find.",
|
| "DATE_TIME is trained as a generic date. DATE_OF_BIRTH stays blocked because the source DATE label carries no birth-date distinction.",
|
| "GOVERNMENT_ID aggregates the source IDCARDNUM and SOCIALNUM labels and is not country-specific; subtype assignment belongs to locale recognizers.",
|
| "SOCIAL_SECURITY_NUMBER is rule-only: 97% of the source SOCIALNUM occurrences are non-US, so the US-specific type cannot be learned from that label."
|
| ]
|
| }
|
|
|
