aditya487/cbi-archive-corpus

Central Bank of Ireland Public Archive Corpus

A page-anchored, provenance-classified corpus of the Central Bank of Ireland's public document archive. 5,568 documents and 89,242 page or pseudo-page rows. PDF rows have true source-page anchors; most Office and archive rows do not.

This is an unofficial derived work. It is not published by, affiliated with, or endorsed by the Central Bank of Ireland.

What makes this different from a pile of scraped PDFs

Two things.

Every page keeps its address. Each row carries the source document's SHA-256, its URL, and its page number in the original PDF. A claim found here can be traced back to a specific page of a specific file whose bytes you can verify. Nothing is a floating snippet.

Every page carries a separately reviewable voice label. The Central Bank hosts two very different kinds of document: material it authored, and submissions that banks, insurers, trade bodies and individuals sent it during public consultations. Those read alike and mean the opposite. A submission from a bank arguing that a rule is burdensome is not a regulatory finding, and treating it as one is the single easiest way to produce confidently wrong analysis from this archive.

v5.1's authorship field collapsed hosting, authorship, document role and institutional voice. v5.2 retains it only for migration and introduces the safe analytical field institutional_voice:

institutional_voice	documents	what it means
cbi-institutional	338	reviewed/rule-supported institutional CBI material
stakeholder	1,739	a respondent writing to the regulator: advocacy or consultation input
external-authority	3	IMF-authored assessments hosted by the Bank
cbi-staff	2	attributable to a named staff member, not automatically the institution
judicial-tribunal	2	a court or tribunal speaking
third-party	2	another evidenced external speaker
mixed	2	a composite; use page-level institutional voice
unknown	3,480	hosted by CBI, but issuer/voice not yet evidenced

unknown is a safety result, not missing data to be folded into the regulator. host, author_org, document_role, voice_review_status, and voice_evidence expose why a row may or may not support a claim about the Bank. See the v5.1 erratum for the 19 confirmed v5.1 speaker errors.

Why the classifier is not a keyword match

An earlier version of this classification treated a consultation-hosted document as a submission only when its filename contained the string response. That missed three whole families:

1. files named cp45-submission-from-aib.pdf
2. files named cp51-feedback-from-generali-paneurope.pdf, which were being filed as Central Bank feedback statements
3. files published under the responder's bare name, such as blackrock.pdf, which carry no cue in the filename at all

Family 3 cannot be resolved from a filename, so classification is two-pass: filename attribution and Central Bank document-type cues first, then the opening pages of the document itself where the filename is silent. Only a narrow set of issuer-specific Central Bank cues takes precedence over stakeholder attribution, which is what keeps note-from-the-financial-regulator-in-relation-to-submissions-received-on-cp43 correctly attributed to the Bank. Generic document-type cues do not: they apply only where the filename carries no attribution, because letting them win is the defect that put AIB's and Bank of Ireland's submissions into the regulator's pile in an earlier version. The rule that produced each label is stored in classification_basis, with a classification_confidence of high, medium or low. 132 regression assertions cover known failure modes. A separate reviewed 32-document regression set includes one mixed composite and two audited submissions outside consultation folders. It was used during development, so it is an error detector, not an independent holdout or a population accuracy estimate. A deterministic 359-document risk queue is included for future independent double review; it is not a gold set or an accuracy sample.

Files

File	Rows	Size	Contents
data/documents.parquet	5,568	1.20 MB	one row per document: URL, hashes, publication evidence, issuer/role/voice, engagement, exact-content cluster and extraction metadata
data/pages.parquet	89,242	46.8 MB	one row per page or pseudo-page, carrying full text and page-level voice evidence
manifests/files.csv.zst	6,984	0.6 MB	the original download manifest: every URL, its SHA-256, bytes, content type, referrers
manifests/conversion-manifest.csv.zst	5,246	0.8 MB	PDF to Markdown conversion record, per document
manifests/conversion-manifest-office.csv.zst	323	varies	conversion details for DOC, DOCX, PPTX and ZIP sources
manifests/provenance-classification.csv.zst	5,568	0.3 MB	full classification audit trail, including the previous label
manifests/extraction-quality.csv.zst	5,568	0.3 MB	per-document extraction fidelity grades
manifests/authorship-overrides.csv.zst	114	varies	SHA-keyed manual voice adjudications
manifests/page-authorship-overrides.csv.zst	4	<1 KB	audited page ranges for the two mixed composites
manifests/conversion-exclusions.csv.zst	1	<1 KB	the one downloaded HTML error body excluded from conversion
manifests/engagement-coverage.csv.zst	182	2 KB	canonical CP/DP identifiers and document-class coverage
manifests/voice-review-scope.csv.zst	359	varies	deterministic high-risk queue for independent voice review
manifests/manifest-build-summary.json	1	<1 KB	hashes and byte sizes for every compressed manifest

The split is deliberate. Most questions need only documents.parquet, which is about 1.2 MB, and never touch the page text at all.

Usage

from datasets import load_dataset

docs = load_dataset("aditya487/cbi-archive-corpus", "documents", split="train")
pages = load_dataset("aditya487/cbi-archive-corpus", "pages", split="train")

Or query it without downloading anything, using DuckDB over HTTPS:

INSTALL httpfs; LOAD httpfs;
SELECT d.title, p.page_number, substr(p.text, 1, 200)
FROM  'https://huggingface.co/datasets/aditya487/cbi-archive-corpus/resolve/main/data/pages.parquet' p
JOIN  'https://huggingface.co/datasets/aditya487/cbi-archive-corpus/resolve/main/data/documents.parquet' d
  USING (document_id)
WHERE p.institutional_voice = 'cbi-institutional'
  AND p.voice_review_status IN ('rule-classified', 'manual-reviewed')
  AND lower(p.text) LIKE '%operational resilience%'
LIMIT 20;

Parquet is columnar, so that query reads only the columns it names.

How the corpus was built

• Crawl snapshot: 25 August 2026, from centralbank.ie, www.centralbank.ie and the open-data portal, using the sitemap plus the CKAN open-data API.
• Downloaded: 6,963 files, 7.476 GB, with 21 URLs failing (all HTTP 404 and all recorded rather than silently dropped).
• Deduplicated by SHA-256: 6,309 unique files. 654 files were byte-identical content served at more than one URL.
• Converted: PDFs through PyMuPDF4LLM with page anchors preserved, selective RapidOCR for scanned documents, and a separate pipeline for the 323 DOCX, DOC, ZIP and PPTX files that a PDF-only filter had originally skipped. OOXML uses a body-order extractor that preserves table position and explicit page breaks; 70 retained legacy DOC conversions state that their historical LibreOffice version was not recorded.
• Validated: every source hash and every output hash recomputed from disk, page markers checked for sequence, orphans checked for. Zero failures.
• Engagement identifiers normalised: cp071, cp-71 and CP71 map to cp71; discussion papers similarly map to dpN. The snapshot contains 149 CP identifiers, 11 DP identifiers and 73 CPs with a proposal, stakeholder responses and a Central Bank feedback statement. The 22 gaps in CP1 to CP171 are published as snapshot coverage gaps, not asserted to be unissued numbers.
• Exact-content clusters retained: 5,493 exact-text clusters cover 5,568 documents; 140 documents belong to a cluster larger than one. Seven are the empty-document cluster, leaving 133 duplicated non-empty documents and 69 excess non-empty records if text is counted once. They remain as separate records because their URLs and publication contexts differ. Use content_cluster_id to avoid double-counting identical text in analysis.

Limitations, stated plainly

• page_basis is not always a page. PDFs carry true source-page anchors. Of the 323 Office/archive records, 137 are a single pseudo-page, 133 use explicit OOXML page breaks, 49 use archive members, two use slides and two have source pages. Only source-page and slide are safe source locations; explicit page breaks are converter boundaries.
• Institutional voice is conservative. 3,480 documents remain unknown because hosting is not issuer evidence. The development set is not an independent holdout, and v5.2 makes no population accuracy claim.
• 87 documents are graded below ok for extraction fidelity: 36 with substantially blank pages, 25 garbled by encoding damage, 19 thin, 7 empty. They are listed in extraction-quality.csv.zst. Chart-heavy statistical releases extract with visible damage even when their numbers survive intact.
• 441,610 characters were recovered in a later pass and are included here. The original converter returned nothing for pages carrying a full-page background image, even where a readable text layer sat underneath, so 1,167 pages were re-extracted directly and a further 258 image-only pages were recovered by OCR. A later targeted OCR pass replaced five garbled CP76 pages with 11,206 readable characters. recovered-pages.csv in the GitHub repository is the cumulative OCR-page ledger; direct text-layer recoveries are reported separately in the extraction review.
• Publication dates are sparse by design. published_at is populated only where source/referrer metadata states an explicit date. analysis_year can additionally use a plausible PDF creation year, but future timestamps are rejected; raw pdf_creation_date is preserved as untrusted metadata.
• ZIP archives are profiled, not transcribed. Most are XBRL taxonomy packages; one held 31,185 schema files. Transcribing them produced 705 MB of Markdown and would have made machine schema dominate every word count. The member listing and namespace profile are included; schema bodies are not.
• Publication volume is not importance. A topic appearing in many documents usually means a consultation drew many responses, not that the topic matters more. Consultation submissions are advocacy and are written to persuade.
• This is a dated snapshot. It cannot contain anything unpublished, unlinked, blocked, or published after 25 August 2026.

Rights, reuse and attribution

This is a mixed-rights corpus, so the machine-readable licence is deliberately other, not blanket CC BY 4.0. Only 71 open-data resources carry explicit CC BY 4.0 metadata. Most Central Bank-authored material is covered by the site's PSI reuse terms, while stakeholder submissions, personal information and other third-party material may fall outside that permission.

Contains Irish Public Sector Information licensed under a Creative Commons Attribution 4.0 International (CC BY 4.0) licence.

Use the attribution above when the PSI licence applies. The CKAN licence title and URL for each open-data resource are retained in manifests/files.csv.zst so exceptions can be checked. The licence excludes personal information, third-party rights, trademarks, crests, logos and official symbols, and prohibits any implication of endorsement.

See RIGHTS-REVIEW.md before redistributing text or using the corpus commercially. This metadata change is conservative and is not legal advice.

Modifications made here: conversion to Markdown and Parquet, deduplication by content hash, provenance classification, extraction-quality grading, and OCR replacement of five pages whose earlier text was 94% decoding noise. The original source bytes remain unchanged in the raw archive.

All generated CSV audit artifacts are UTF-8 without a byte-order mark and use LF line endings, so their locked hashes are stable across Windows and Unix.