{
  "status": "PRELIMINARY",
  "release_date": "2026-10-22",
  "dataset": "banking77",
  "source_rows": 13083,
  "rows_in_scope_before_cleaning": 13083,
  "rows_with_redaction_before_deduplication": 2317,
  "retained_rows": 13006,
  "retained_rows_with_redaction": 2302,
  "excluded_normalized_duplicates": 68,
  "excluded_conflicting_label_rows": 8,
  "excluded_cross_split_duplicates": 32,
  "excluded_no_usable_text": 1,
  "redacted_span_counts_before_deduplication": {
    "AMOUNT": 197,
    "PET_NAME": 3,
    "DATE": 608,
    "ORGANIZATION": 1196,
    "TIME": 177,
    "GROUP": 63,
    "LOCATION": 416,
    "PERSON": 105
  },
  "splits": {
    "test": 3073,
    "train": 9933
  },
  "intent_count": 77,
  "checks": {
    "count_reconciliation": true,
    "no_normalized_duplicate_text": true,
    "no_cross_split_normalized_duplicates": true,
    "identifier_pattern_scan_passed": true
  },
  "privacy_method": "Regular-expression identifier/contact/address screening and spaCy en_core_web_sm 3.8.0 named-entity redaction. Automated screening is not a guarantee of anonymity.",
  "deduplication": "NFKC, casefold, retain alphanumeric characters only. Identical normalized texts are deduplicated across all three released customer-service datasets, with evaluation preferred over training and BANKING77 then CLINC150 then Bitext as source tie-breaker. Conflicting labels for identical normalized text within a source are excluded.",
  "source_files": [
    {
      "file": "banking-train.csv",
      "url": "https://raw.githubusercontent.com/PolyAI-LDN/task-specific-datasets/master/banking_data/train.csv",
      "bytes": 839073,
      "sha256": "b06e26ac675513959a63135f11b94ea7786ed02da65db93a5650d8838cbc664b"
    },
    {
      "file": "banking-test.csv",
      "url": "https://raw.githubusercontent.com/PolyAI-LDN/task-specific-datasets/master/banking_data/test.csv",
      "bytes": 239961,
      "sha256": "d12d6e3bc4c3103966ae786dc435913c0c563dfa328f5a3646d0e62cfeeb474d"
    },
    {
      "file": "banking-license.txt",
      "url": "https://raw.githubusercontent.com/PolyAI-LDN/task-specific-datasets/master/LICENSE",
      "bytes": 18650,
      "sha256": "7e7170e3cebf88a9f60c7b8421418323c09304da1af4d5e90f4da1dc1c8a2661"
    },
    {
      "file": "banking-readme.md",
      "url": "https://raw.githubusercontent.com/PolyAI-LDN/task-specific-datasets/master/README.md",
      "bytes": 4735,
      "sha256": "7b9ce9069931cc2de89dce8c7ffb53d07806cf413601ccf133b9af5227b3a373"
    }
  ],
  "independent_privacy_revision": {
    "date": "2026-10-23",
    "version": "2026-10-23-privacy-v1",
    "records_rescreened": 13006,
    "records_with_additional_location_redaction": 5,
    "additional_location_spans": 5,
    "rows_removed": 0,
    "records_after": 13006,
    "record_ids_recomputed_for_changed_text": true,
    "source_row_mapping_included": false,
    "method": "Independent regex, named-entity and capitalization-pattern review of the complete released text. Five residual fine-location spans manually confirmed and generalized to LOCATION placeholders. False positive currency codes, ATM/PIN terms and historical attribution were retained.",
    "limitations": "This targeted review reduces exposed details but does not certify anonymity. Matching against public upstream text can remain possible; unidentified names may escape automated screening."
  }
}