About the Project

Dataset-Maker creates reproducible torn-document datasets with the answer key built in.

Upload a PDF, choose how many fragments you want and how rough the tears should be, and Dataset-Maker splits each page into non-overlapping pieces while recording exactly where every fragment belongs and which pieces share a torn edge.

The goal is simple: document-reassembly research needs torn documents where the correct solution is known. Physical tears do not come with ground truth, and ad-hoc generators are hard to compare or reproduce.

What it guarantees

• Strict partitioning: every page pixel belongs to exactly one fragment, with no overlaps or gaps. Every generated page is verified before export.

• Deterministic tears: the same seed, tear parameters and DPI reproduce the same partition, placement and adjacency.

• Exact ground truth: every fragment's page position and dimensions, plus the list of fragment pairs that share a torn edge.

• Reproducible provenance: `manifest.json` records seeds, generation parameters, the input SHA-256 and relevant library versions.

• Inspectable output: the web app lets you inspect the source page, partition map, adjacency graph, fragments and manifest before you download the export.

The tearing model starts from a Voronoi partition and domain-warps the coordinate field with value noise, creating irregular boundaries while preserving the strict partition guarantee.

dm-bench

Dataset-Maker also ships dm-bench 0.1.0, a versioned benchmark for torn-document reassembly with easy, medium and hard tiers, public val and test-dev splits, and a held-out test split generated from a private secret, with a published commitment so its answers cannot be changed later.

It includes an evaluation harness for placement accuracy, adjacency F1, perfect pages and candidate ranking, plus [email protected], a deterministic classical baseline that reads only the puzzle fragments.

On the held-out test split, the baseline reaches 0.993 direct accuracy on easy, 0.865 on medium and 0.246 on hard. It solves 0% of hard pages perfectly, despite Hit@1 remaining above 0.9, leaving a clear gap between identifying neighbouring pieces and reconstructing the full page.

Benchmark pages are synthetic and generated from a seed, so there is no personal or copyrighted data; a real-scan set is future work.

dm-bench 0.1.0 is published on Hugging Face with a DOI, and Dataset-Maker is open source under the MIT License.