BayKolm/pdf-to-markdown-test-corpus

PDF-to-Markdown Test Corpus

18 small PDFs, each built to break a PDF-to-Markdown converter in one specific way, plus the measured output of one converter against all of them.

If you are writing a converter, or choosing one, the hard part is not the happy path. It is knowing what happens when a document has two columns, or a heading that is only bold, or a scanned page in the middle. This corpus is meant to make that testable in about a minute, and to give you somewhere to point when you find a bug.

git clone https://huggingface.co/datasets/BayKolm/pdf-to-markdown-test-corpus
cd pdf-to-markdown-test-corpus
python3 verify.py          # structural check + text extraction, no dependencies

Then open your converter, drop in the files from tests/, and compare against results.json.

What is in here

File	What it isolates
01-single-column.pdf	The case everything is built for — a baseline to compare against
02-two-column-shared-baseline.pdf	Two columns on identical baselines
03-table.pdf	Table cells sharing a baseline
04-no-text-layer.pdf	A page with no text operators at all
05-two-column-offset.pdf	Two columns that overlap for only part of the page
06-three-column.pdf	Three columns sharing baselines
07-bold-headings-body-size.pdf	Headings marked out only by weight, not size
08-running-header-2-pages.pdf	The same header and footer repeated on every page
09-hyphenated-breaks.pdf	Words split across lines with a trailing hyphen
10-toc-wide-gap.pdf	A wide horizontal gap inside one visual line
11-scan-styled.pdf	Bars standing in for a scanned page
12-mixed-worst-case.pdf	Title, two-column paragraph and table on one page
13-heading-thresholds.pdf	One line at each of six font sizes around the usual thresholds
14a-row-major.pdf	Identical geometry to 14b, text written row by row
14b-column-major.pdf	Identical geometry to 14a, text written column by column
15-mixed-first-page-text.pdf	Page 1 has text, page 2 is a scan
16-mixed-middle-page-text.pdf	Scan, text, scan
17-compressed-stream.pdf	A FlateDecode-compressed content stream

Two rules apply to all of them. Each file isolates one behaviour, so a surprising result points at a single cause. And where two files are compared they differ in one variable only — 14a/14b are the clearest case: same page size, same font, same coordinates, the same four left-column and four right-column strings, and the same nine text items in total. The only difference is the order those text operators appear in the content stream. (Each file does carry its own heading, Row Major and Column Major, so that you can tell at a glance which one produced a given result; that heading is the one string that differs.)

The pair worth looking at first

14a-row-major.pdf and 14b-column-major.pdf are visually indistinguishable. Open them side by side and you cannot tell which is which.

Yet one converter (the one whose results are in results.json) produced:

Left column, line one. Right column, line one.

for 14a, and

Left column, line one.Right column, line one.

for 14b — no space at all between the columns.

Both are wrong, but the second is far more dangerous, because line one.Right column reads as a typo rather than as two columns that have been fused together. Nobody proofreading the output catches that.

The cause is not the geometry. It is that PDF.js emits the spacing between two text runs as its own synthetic text item when the runs are written consecutively, and omits it otherwise. Same page, same coordinates, different outcome — a property of the file's internal byte order that no amount of looking at the page will reveal.

Measured results

results.json holds the verbatim output of one converter against every file, measured 2026-09-17. Structure:

{
  "converter": { "name": "...", "engine": "...", "measured": "..." },
  "results": [
    {
      "file": "01-single-column.pdf",
      "intent": "what the file is built to test",
      "status": "the string the tool reported",
      "output": "verbatim Markdown, or null if it errored"
    }
  ]
}

Your converter's results will differ. That is the point — the value is in having a fixed set of inputs you can compare over time, and in the entries like heading_measured in 13-heading-thresholds.pdf, which records which font sizes became which heading levels.

Rebuilding the corpus

python3 make_pdfs.py    # regenerates tests/ — byte-identical when run twice
python3 verify.py       # checks xref offsets, /Length, page and stream references

No dependencies. Both scripts are standard library only, and the PDFs are written byte by byte rather than generated with a library, so you can read the exact content stream behind any result. When a converter does something surprising it matters that the input is not a black box.

verify.py is deliberately not a PDF parser. It checks the invariants a reader depends on — that startxref points at the xref table, that every xref entry points at a real object header, that declared /Length matches the actual stream length, that every page's /Contents resolves — and then pulls the text back out with a regular expression. If it passes, the files are well-formed enough for a conforming parser.

Contributing

A new file is welcome if it isolates one behaviour and comes with its measured output. The most useful contributions are the ones that are currently not covered: forms and annotations, rotated pages, right-to-left text, embedded fonts with unusual encodings, incremental updates, and PDFs from the awkward corners of the format.

Please keep the naming convention NN-what-it-tests.pdf, keep the file small (under about 5 KB), and add the entry to results.json.

Mirror

The canonical source for this corpus is the Git repository GordonGuoJin/pdf-to-markdown-test-corpus. This dataset is a mirror of it, published here so that the corpus can be found and cited alongside other document-processing datasets. Changes land in the Git repository first.

Licence

The PDFs and scripts are released under CC0 1.0 — public domain. Use them for anything, including commercially, with no attribution required.

They are synthetic documents built for testing and contain no third-party content.

---

Built while testing pdfmdtools.com, a browser-side PDF-to-Markdown converter, and published so the test cases do not have to be rebuilt by everyone who needs them.