aditya487/cbi-archive-raw

Central Bank of Ireland Archive: original source files

6,309 original files, 6.56 GB. Every PDF, spreadsheet, Word document and archive gathered from the Central Bank of Ireland's public website, stored by content hash so that a search result can be turned back into the document a human would actually read.

This is the raw tier. If you want the text, you almost certainly want aditya487/cbi-archive-corpus instead: 5,568 documents and 89,242 page or pseudo-page rows as Parquet, queryable over HTTPS without downloading anything.

Come here when the text is not enough: charts, diagrams, scanned pages, tables that did not survive extraction, and the 87 documents graded below ok.

How files are addressed

Every file lives at its own SHA-256:

<first 2 hex chars>/<next 2>/<full sha256><extension>

So ae0acdc7...b49e.pdf is at ae/0a/ae0acdc7...b49e.pdf. The identifier and the location are the same string, which means no lookup service is needed to get from a search hit to the original:

search the text  ->  source_sha256  ->  ae/0a/ae0acdc7....pdf

source_sha256 appears on every row of documents.parquet in the corpus dataset, so the two repositories join without any shared infrastructure.

Note for anyone reading older documentation: files are at the repository root, not under a blobs/ prefix. Some early drafts said otherwise and were wrong.

What is here

Format	Files
PDF	5,246
XLSX	447
CSV	246
DOCX	201
DOC	71
ZIP	49
XLS	42
XML	5
PPTX	2

These are URL/path-declared formats. Magic-byte checks found three .pdf paths containing Word bytes, one containing HTML, and two .docx paths containing PDF bytes. The raw objects are preserved unchanged; converters dispatch on detected content where necessary.

Plus three metadata files:

• blob-catalog.csv maps every hash to its path, format, byte count, and every URL that served it, including aliases. 6,309 rows.
• blob-summary.json records how the tree was built, including the 654 duplicate copies collapsed by content addressing, which saved 917 MB.
• page-catalog.csv maps immutable HTML page-context snapshots to source URLs. It is empty for the August 2026 snapshot because the original crawler retained referrer URLs/statuses but not HTML bodies. Future refreshed pages are stored under page-context/<sha-prefix>/<sha256>.html by default.

Getting a file

python get_source.py --search "operational resilience" --limit 3 --fetch

get_source.py lives in the GitHub repository. It searches the Parquet text, resolves the hash, downloads the original, and verifies the SHA-256 before writing it. Or fetch directly:

https://huggingface.co/datasets/aditya487/cbi-archive-raw/resolve/main/ae/0a/<sha256>.pdf

For a reproducible citation, replace main with the immutable raw_revision recorded in RELEASE.lock.json in the GitHub repository.

The Hub exposes SHA-256 metadata for the 5,296 files stored through LFS. The remaining 1,013 are ordinary Git blobs, for which the tree API proves path and size but does not expose the file's SHA-256. blob-catalog.csv carries the expected hash for all 6,309; publish/verify_dataset.py checks every pinned path and size and every available LFS hash.

Rights and reuse

Read this before redistributing.

These documents were published by the Central Bank of Ireland on its public website. They are gathered here for research. Copyright and database rights remain with the Central Bank of Ireland and, where applicable, with the third parties who wrote the submissions.

• 71 files carry an explicit Creative Commons Attribution 4.0 licence, from the Bank's open-data portal.
• 6,238 files carry no explicit licence. Their reuse rests on the Central Bank's website terms, which exclude personal information and third-party rights from the general permission to reuse.

That exclusion is the reason this repository is framed as research material rather than as an open dataset. It is a mirror of public documents, offered so that findings drawn from the text can be checked against the originals.

Personal data

Roughly a third of these documents are consultation submissions, most from banks, insurers, law firms and trade bodies. A small number are from private individuals writing in a personal capacity.

The corpus has been screened for personal data. Findings, in full:

• No national identifiers. All 17 candidates matching the Irish PPS number pattern turned out to be company VAT numbers, which share the same shape (seven digits and a letter). Every one appeared on a corporate letterhead.
• No personal payment details. All 4 IBAN matches are the Central Bank's own published accounts for levy and fee payments.
• 18 candidates were manually reviewed: nine are private individuals and nine are people writing in a public or professional role. Twelve of the 18 are stakeholder documents. The nine private submissions are 0.52% of the 1,739 stakeholder documents in v5.2. The decisions are recorded in RIGHTS-REVIEW.md and qa/individual-submission-review.csv.
• Email addresses and phone numbers appear widely but are overwhelmingly corporate contact points already printed on letterheads.
• The Central Bank applies its own redaction upstream: 206 documents reference redaction, and some personal email addresses arrive already masked.

The full method and results are in RIGHTS-REVIEW.md in the GitHub repository.

Takedown

If you are named in one of these documents and would prefer not to be, the material will be removed. No justification is required and no argument will be made.

Open a discussion on this dataset, or contact the maintainer through the GitHub repository. Please include the SHA-256 or the URL if you have it, though a description is enough.

The same applies to any rights holder who objects to a document being mirrored here.

What this is not

• Not official. Not affiliated with, endorsed by, or connected to the Central Bank of Ireland.
• Not authoritative. For the current text of any regulation, rule or guidance, go to centralbank.ie. Documents here are a snapshot taken in August 2026 and some are many years old and superseded.
• Not complete. It is what a crawl of the public site found, not the Bank's full record.
• Not advice. Nothing here is legal, financial or regulatory advice.

One thing that will mislead you if you skip it

The corpus contains two kinds of document that read alike and mean the opposite: the regulator setting expectations, and firms lobbying the regulator about them. A consultation response from a bank arguing that a requirement is disproportionate is advocacy, not a regulatory finding, and treating it as one is the easiest way to produce confidently wrong analysis from this data.

Use institutional_voice, not the legacy authorship field, for this decision. v5.2 labels 338 documents cbi-institutional, 1,739 stakeholder, three external-authority, two each cbi-staff, judicial-tribunal, third-party and mixed, and 3,480 unknown. The large unknown class is deliberate: a file being hosted on centralbank.ie is not proof that the Bank authored it. Check voice_review_status and the page-level fields before quoting.

Citation

@misc{cbi_archive_raw_2026,
  title  = {Central Bank of Ireland Archive: Original Source Files},
  author = {Prakash, Aditya},
  year   = {2026},
  url    = {https://huggingface.co/datasets/aditya487/cbi-archive-raw}
}