pdf pdf |
|---|
tuandunghcmut/multipage-dla-pdf-corpus-full
Generated at: 2026-05-02T16:24:49+00:00
Scope
Public-safe PDF corpus export from the multipage DLA pdf_preparation subproject. The export is filtered to accepted records with redistribution=yes, without policy-denied URLs, without Mixed-Verify records, without training-only records, and without records that require PII review. The corpus is designed for research on multipage Document Layout Analysis (DLA), especially layouts where headings, tables, forms, legal text, annexes, lists, and cross-page continuations appear in realistic applied documents.
This dataset is a public-safe corpus snapshot for research and engineering review. It is not reviewed gold data, and it should not be cited as a final benchmark without additional review.
Export policy
- Catalog source:
metadata/catalog.jsonl - Status filter:
accepted - Include Mixed-Verify:
False - Public-safe filter:
True - Repository visibility: public
- Local filesystem paths are not part of the public dataset contract. Use
hf_repo_pathfor files inside this repository andsource_urlfor provenance.
Aggregate
| Metric | Value |
|---|---|
| PDFs | 1814 |
| Pages | 76140 |
| Source modules represented | 33 |
Family taxonomy
The A-E labels are domain families used to balance the PDF corpus. They are not model labels and are not intended as document-layout categories.
| Family | Meaning | PDFs | Pages | Example source modules |
|---|---|---|---|---|
| A | Standards, cybersecurity, quality, and ISO-pattern documents. | 552 | 17022 | ietf_rfcs, nist_sp800_curated, fips_publications, epa_quality_guidance_curated |
| B | Financial, banking, consumer-finance, supervision, and institutional reports. | 103 | 5465 | fincen_guidance, federal_reserve_pubs, ffiec_it_handbooks, fed_annual_reports_curated |
| C | Legal, regulatory, government audit, rule, bill, and court-opinion documents. | 152 | 3217 | federal_register_articles, supreme_court_opinions, congress_bills_curated, ftc_reports |
| D | Forms, tax instructions, procurement, compliance, and reporting documents. | 106 | 3034 | irs_publications, irs_current_forms_landing, irs_forms_curated, uscis_forms_instructions |
| E | Common/everyday multipage public documents used to preserve general layout coverage. | 901 | 47402 | archive_org_pd_pdfs, gutenberg_pdf_books, va_benefits_handbooks, national_archives_publications |
Legal, redistribution, and privacy posture
This dataset is curated as a lawful, public-source research corpus for multipage document layout analysis. The collection process is designed to avoid private, confidential, paywalled, credential-gated, or access-restricted documents.
Key safeguards:
- Source scope: crawlers target public institutional, government, public-domain, standards, regulatory, and public-information PDF sources.
- Deny-list policy: known restricted or paid standards sources are blocked by deterministic source policy and are not intentionally included.
- License policy: each source is assigned a deterministic
license_tagandredistributionpolicy; reviewers can audit these fields per PDF inmetadata/catalog.jsonl. - Privacy policy: the pipeline does not crawl private drives, emails, chats, internal company repositories, personal cloud folders, or user-submitted private documents.
- PII safeguards: sources that may require additional privacy review are marked with
pii_review_required; such records should not be promoted beyond private/internal research use without review. - Local-path privacy: local filesystem paths are not exposed as dataset metadata; the Hub catalog uses
hf_repo_path,source_url, hashes, sizes, and redacted audit placeholders. - Review status: records marked
review_requiredorMixed-Verifyremain explicitly flagged for additional legal/privacy review before any broader public redistribution.
Within these controls, the dataset is intended to be a privacy-respecting and license-aware research corpus. The metadata is structured so reviewers can trace every file back to source policy, provenance, and quality status. This data card is not legal advice; final release decisions should still follow the project's documented review process.
Source and provenance model
Each PDF is traceable through metadata/catalog.jsonl. The catalog has one JSON object per PDF. Reviewers can inspect:
source_idandsource_display_nameto identify the crawler/source definition.source_urlandsource_domainto identify where the PDF came from.hf_repo_pathto locate the file inside this Hugging Face repository.sha256,file_size_bytes, andpage_countto support reproducibility checks.license_tag,redistribution,status, andreview_statusto understand release policy and data quality status.
The data card summarizes the corpus. The catalog is the authoritative per-file audit table.
Largest represented source modules
| Source ID | Display name | Family | PDFs | Main source domains |
|---|---|---|---|---|
archive_org_pd_pdfs |
Internet Archive PD multi-page PDFs | E | 814 | archive.org |
ietf_rfcs |
IETF RFC archive | A | 500 | www.rfc-editor.org |
federal_register_articles |
Federal Register article-level PDFs | C | 81 | www.govinfo.gov |
gutenberg_pdf_books |
Project Gutenberg PDF books | E | 73 | www.gutenberg.org |
fincen_guidance |
FinCEN advisories, bulletins, and fact sheets | B | 49 | www.fincen.gov |
federal_reserve_pubs |
Federal Reserve Board publications | B | 32 | www.federalreserve.gov |
supreme_court_opinions |
U.S. Supreme Court slip opinions | C | 30 | www.supremecourt.gov |
irs_publications |
IRS publications and instructions | D | 29 | www.irs.gov |
nist_sp800_curated |
NIST SP-800 cybersecurity (curated) | A | 27 | nvlpubs.nist.gov |
congress_bills_curated |
Congress.gov bills (curated recent) | C | 26 | www.govinfo.gov |
irs_current_forms_landing |
IRS current forms and instructions landing page | D | 23 | www.irs.gov |
irs_forms_curated |
IRS form instructions (curated) | D | 17 | www.irs.gov |
fips_publications |
NIST FIPS standards | A | 12 | nvlpubs.nist.gov |
uscis_forms_instructions |
USCIS forms and instructions | D | 12 | www.uscis.gov |
ffiec_it_handbooks |
FFIEC IT examination handbooks | B | 10 | ithandbook.ffiec.gov |
irs_instructions_boost |
IRS additional form instructions curated | D | 10 | www.irs.gov |
uscis_forms_boost |
USCIS additional forms and instructions curated | D | 10 | www.uscis.gov |
ftc_reports |
FTC consumer-protection reports | C | 7 | www.ftc.gov |
uscourts_rules |
U.S. Courts rules of practice and procedure | C | 7 | www.uscourts.gov |
va_benefits_handbooks |
VA veterans benefits handbooks | E | 7 | www.vba.va.gov, benefits.va.gov |
Metadata schema
The uploaded catalog is intended to be readable by both scripts and reviewers. Important fields are:
| Field | Meaning |
|---|---|
catalog_id |
Stable row identifier for the exported catalog. |
document_id |
Document identifier from the source manifest. |
source_id |
Crawler/source module that produced the record. |
source_display_name |
Human-readable source name. |
family |
A-E domain family used for balancing and reporting. |
family_name |
Longer family label when available. |
source_url |
Original URL from which the PDF was discovered or downloaded. |
source_domain |
Domain extracted from the source URL for provenance review. |
hf_repo_path |
Path of the PDF inside this Hugging Face dataset repository. |
status |
Filtering status such as accepted, quarantined, or pending, depending on the export filter. |
license_tag |
License/policy tag assigned by deterministic source policy. |
redistribution |
Whether the current policy allows redistribution in this private export. |
sha256 |
Content hash for deduplication and reproducibility checks. |
file_size_bytes |
PDF size in bytes. |
page_count |
Detected page count used by corpus filters. |
language |
Detected or source-assumed language. |
quality_score |
Heuristic score for prioritization, not a gold-quality label. |
quality_features |
Supporting quality signals used by the filter/scoring pass. |
is_scanned_estimate |
Heuristic estimate for scanned/image-heavy documents. |
review_status |
Data quality status such as structurally_validated or review_required. |
pii_review_required |
Flag for records that may need extra review before broader release. |
raw_path_redacted |
Redacted local audit placeholder. Use hf_repo_path for dataset access. |
By status
| Status | PDFs |
|---|---|
| accepted | 1814 |
By license tag
| License tag | Meaning | PDFs |
|---|---|---|
| PD | Public-domain or open public-domain-like source policy. | 887 |
| Free-Standard-IETF | Freely accessible standards/RFC-style source policy. | 500 |
| PD-US-Gov | United States government/public-sector source policy used for redistribution review. | 422 |
| Free-Standard-ECMA-OASIS-OGC | Source-specific license/policy tag. | 5 |
Important limitations
- Per-file metadata is authoritative in
metadata/catalog.jsonl. Mixed-Verifyfiles need license review before public redistribution.- This private dataset may include candidate annotations and should not be described as reviewed gold data.
quality_scoreis a heuristic corpus-management score, not a human rating.- The current snapshot reflects the crawler/filter state at export time. Future exports may add sources or rebalance families.
- Downloads last month
- 25