dm-bench 0.1.0

Torn-document reassembly benchmark. Synthetic, seeded, licence-clean pages are torn into non-overlapping fragments, and a solver must place every fragment back on the page with a rigid pose. Full benchmark card, metrics and baseline: docs/BENCHMARK.md. Code: Arittra-Bag/Dataset-Maker.

Contents

tier	val pages	test-dev pages	test pages
easy	39	13	48
medium	37	16	50
hard	33	19	52

• puzzles/<tier>/<split>/<page_id>/: puzzle.json plus fragment PNGs (RGBA).
• answers/<tier>/{val,test-dev}/: ground-truth poses and adjacency.
• test answers are held out. They are derived from a maintainer secret, and the commitment is in benchmark.json (test_commitment), see below.
• SHA256SUMS (bytes) and CONTENT.sha256 (decoded content) cover every file. release_sha256 for this release: 6fe00ba8fac1e39b62cfa78a266095ba7e24c594d0fb2b5442b1148266b185ed.

The train split is not shipped. Regenerate it with python -m src.bench build --splits train from the code repository.

Use

python -m src.bench verify <this folder>
python -m src.bench solve --release <this folder> --out solutions
python -m src.bench eval  --release <this folder> --solutions solutions --splits val test-dev

How to use the splits

• Fit learned methods on train (built on demand), tune on val.
• Use test-dev for public development checks.
• The held-out test split is scored through the maintainer, since scoring needs the secret. There is no submission server.

Held-out test: commitment and reveal

• secret_sha256: 74d2bc987213c9e796bbf4de54dfad09caf02e40cfee842ce05533643f2adc5d
• answers_sha256: bdc3a674a6ca064cf029f8d6dc37825b951daf9703c1394afc289d5a8360ec9c

The v0.1 test secret is published verbatim in this dataset repository when the v0.2 held-out test split is published, or on 2027-09-30, whichever comes first. Until then nobody outside can verify the commitment. Each answer record carries its page id, tier, split, document, page and fragment ids, so adding or dropping a test page changes answers_sha256. The exact hashing recipe, byte for byte with a standard-library Python snippet, is in Commitment recipe.

Rendering

Geometry deterministic, pixels renderer-dependent. Tears, poses and adjacency depend only on the seed, the tier parameters and the page size. Fragment pixels, the per-fragment ink_frac in each answer and the blank counts in benchmark.json come from the PyMuPDF render. This release was built with PyMuPDF 1.24.10 (MuPDF 1.24.9), recorded in build_env. Rebuild with those versions to reproduce it exactly. How much another PyMuPDF version changes pixels and scores is measured in the code repository (issue #10).

Baseline (edge-greedy@0.1, eval_version 1.0, held-out test)

Read the hard tier first. Hit@1 is 0.928 over all 52 pages: on an average page, 92.8% of the fragments that have a ground-truth neighbour get one ranked first. Yet the baseline reassembles 0 of 52 pages perfectly, where perfect means every fragment within 1% of page width under one rigid alignment.

tier	perfect pages	Hit@1	direct_acc [95% CI]	neighbor_acc
easy	0.958	0.997	0.993 [0.981, 1.000]	0.994
medium	0.300	0.959	0.865 [0.820, 0.907]	0.851
hard	0.000	0.928	0.246 [0.224, 0.267]	0.640

Source: docs/results/dm-bench-0.1.0_edge-greedy-0.1_eval-1.0.json in the code repository.

Citation

DOI 10.57967/hf/10638. BibTeX is in the code repository README.