Genuine PDF Corpus (PDF-Forensics)

Issued or published, unmodified original; not an edited derivative, synthetic document, specimen, blank form or reference document.

Each file record includes SHA-256, byte count, relative package path, source provenance (the source and, where recorded, its URL), document labels and licence information. Null fields are unknown, not negative findings. There is no edited-pair or edit-operation catalogue in this version.

The Genuine PDF Corpus (PDF-Forensics) collects original PDFs from public issuers and public datasets. Its address keeps “1m” from earlier versions so that links stay stable; version 1.2.0 contains 82,955 original PDFs, all redistributed as files.

Version 1.2.0 keeps only the files that our Trustable detection engine, release 0.18.0 (build 83416ce4, structural model v10j_b), read CLEAN in a full scan on 31 October and 1 November 2026: 82,955 of 95,245 candidates. The other 12,290 were withheld: files it flagged (SUSPICIOUS or FORGED), files it could not reach a conclusion on or did not finish scanning, and files we hold out for our own evaluation. DATASET_CARD.md and summary.json record the same totals. Link-only records (files listed by address but not shipped) were never scanned and are not part of this version.

Internal employees’ personal files and customer-provided files are excluded. A source that is held for unresolved permission, privacy, attribution or provenance is not silently treated as approved for release.

The genuine label describes source provenance and the absence of a recorded edit. It does not certify the issuer, the contents, or the absence of a change before collection. 0 shipped files have a recorded human-reviewed label.

Document type, language and PDF format are machine-assigned or source-provided labels. Missing and undetermined values remain visible. Source and format imbalance can affect a detector’s alert rate.

The CLEAN screen is a selection, not a label audit (see “Read the labels in context” above). A later engine version may flag some of these files.

Report input count, completed checks, unsupported inputs and failures alongside any detector score. Give the alert rate on the files a detector actually assessed, and call it a false-positive rate only once you have independent ground truth.

Some source PDFs are public court records. The release’s existing screen checks for Social Security numbers and full account numbers; it does not remove names or home addresses. Public-source status is not a promise of de-identification.

Do not use these records to target or profile named people. Check each record’s source conditions and the data card before use. Report a privacy concern with the file hash through the takedown contact below.

The package contains 6 independent PDF archives and one metadata archive. A part is a storage group, not a train/test split.

Start with the metadata: read DATASET_CARD.md, DOWNLOADS.md, LICENSES.md, ATTRIBUTION.md, summary.json, licences/ before downloading the PDF parts. For a partial download, verify only the matching checksum lines; files you have not downloaded are not verified.

# Verify every downloaded archive against its published checksum.
sha256sum -c ARCHIVE_SHA256SUMS

# Extract separate parts into one new directory; do not concatenate them.
mkdir dataset
tar --zstd --extract --file metadata.tar.zst --directory dataset --no-same-owner
tar --zstd --extract --file part-00000.tar.zst --directory dataset --no-same-owner

cd dataset
sha256sum -c SHA256SUMS

# Verify the PDF bytes in one extracted part.
python3 -c 'import json, sys
for line in open(sys.argv[1]):
    r = json.loads(line)
    print(r["sha256"] + "  " + r["path"])' parts/part-00000.jsonl |
  sha256sum -c --quiet

Every original retains its source licence or recorded legal basis. The 82,955 shipped PDFs cover 33 licence records, in three groups that do not overlap: 49,246 files under a licence that permits redistribution, 26,914 under one that permits it with attribution, and 6,795 public court records for which no licence statement was found, shipped under our rule for public court records.

There is no blanket licence replacing these source conditions. Keep notices, required credit and any per-file use restrictions. The source records, full texts and credits are in LICENSES.md, ATTRIBUTION.md and licences/.

Do not misrepresent a research copy as an issuer’s authenticated document, use it to deceive, or target a person named in a source. A public URL does not remove privacy responsibilities.