README.md

metadata

license: mit
task_categories:
  - tabular-classification
tags:
  - pdf
  - pdf-repair
  - test-corpus
  - synthetic
pretty_name: PDF Corruption Lab
size_categories:
  - 1K<n<10K
configs:
  - config_name: features
    data_files: features/train.csv
    default: true

pdf-corruption-lab

A fully synthetic corpus of damaged PDF files with a known kind and severity of damage, the scripts that generate it, and a ready-to-use feature table. It is meant for testing and benchmarking PDF repair / recovery tools and for training small damage-type classifiers.

Everything here is original: documents are built from scratch with pdf-lib (MIT). No third-party documents or models are included.

What is in this repository

Path	Description
output/	7 hand-checkable reference files: one valid PDF and 6 damaged variants, plus manifest.json
training/	1,200 randomized PDFs (200 per class: pages, fonts, text and images vary) and labels.csv
features/train.csv	12 byte-level structural features + label for each training file (powers the dataset viewer)
generate.js	regenerates the reference files in output/
gen_training.js	regenerates the randomized training set (seeded, reproducible)

Damage classes

Class	Damage	Severity (1-5)
valid	none	0
missing_eof	trailing %%EOF removed	1
invalid_startxref	startxref value is non-numeric	2
scrambled_xref	xref offsets replaced with bogus values	3
bad_root_reference	trailer /Root points to a missing object	3
truncated	file cut at a random point (30-90 % of its length)	4-5

Features (features/train.csv)

has_pdf_header, eof_in_last_1k, ends_with_eof, startxref_present, startxref_numeric, startxref_points_to_xref, xref_valid_fraction, has_trailer, root_object_exists, obj_endobj_gap, stream_endstream_gap, log10_size, and label. The extraction code is in the model repo: utilvo-com/pdf-damage-classifier.

Regenerate

npm install
node generate.js ./output
node gen_training.js ./training 200

Suggested evaluation for repair tools

For each damaged file, run your tool and record:

1. Does the output open in a strict parser?
2. How many pages were recovered?
3. Was the embedded image preserved (where present)?

Limitations

All files are generated by pdf-lib with classic xref tables (no object streams or xref streams) and each file has a single kind of damage. Real-world corruption (bit flips, mixed damage, files from other producers, encryption) is not represented. Treat results here as a controlled benchmark, not as a guarantee for real files.

Disclosure

Maintained by the Utilvo team. The organization profile links to our website.

License

MIT. See LICENSE.