docling-project/regression-dataset-for-docling-parse

Regression Dataset for `docling-parse`

This repository contains the reference dataset used as a regression test corpus for `docling-parse`.

Its purpose is to make parser and renderer changes safe: when behavior changes in `docling-parse`, the test suite can compare the current output against the expected artifacts stored in this dataset.

Correct workflow to add new files

cp /path/to/new.pdf regression/new.pdf

git add regression/new.pdf
git lfs status
git commit -s -m "Add new regression PDF"

git lfs push origin main
git push origin main

What this dataset is used for

The dataset serves two related parts of the `docling-parse` test suite:

• —parse: PDF parsing outputs are checked against stored ground-truth structures, extracted content, and document-specific regression fixtures.
• —renderer: rendering outputs are checked against stored render instructions, bitmap artifacts, and page images.

Reference test code is included in this repository under _docling_parse/_tests, in particular:

• —_docling_parse/_tests/test_parse.py
• —_docling_parse/_tests/test_renderer.py

Repository contents

The dataset is organized into a few main groups:

• —regression/: source PDF files used for regression coverage.
• —groundtruth/: expected parse outputs for selected pages and documents.
• —groundtruth_renderer/: expected renderer outputs such as instruction JSON, bitmap metadata, exported bitmap files, and full-page images.
• —cases/, errors/, synthetic/: additional fixtures covering focused edge cases, failure scenarios, and synthetic test inputs.

Why this exists

PDF parsing and rendering are both sensitive to small implementation changes. This dataset helps detect unintended regressions in:

• —text extraction
• —layout and geometry
• —annotations, forms, and shapes
• —bitmap extraction
• —page rendering instructions

In short, this repository is the regression baseline for both the parse and the renderer parts of `docling-parse`.