docatlas-pl-clean-v1

Deduplicated, screened Polish subsets of DocAtlas Instruct: 1714 pages total across four acquisition batches (500 × 4 raw rows). Root files describe batch 1; batches/batch2-20261017/, batches/batch3-20261017/ and batches/batch4-20261017/ are self-contained.

Provenance

• Upstream dataset: ahmedheakl/docatlas_instruct, revision bff2e5bddb3c5f9625029d45e41b2ad75268ea22 (head at acquisition, last modified 2026-10-02), Polish-language rows only.
• Intermediate snapshot: docatlas-pl-candidate-500, same revision.
• Cleaning script: prepare_clean_corpus.py (deterministic, rerunnable against the source snapshot).

Batches

	batch 1 (root)	batch 2 (batches/batch2-20261017/)	batch 3 (batches/batch3-20261017/)	batch 4 (batches/batch4-20261017/)
acquired	2026-10-14, source rows 133–15489	2026-10-17, source rows 15548–37991	2026-10-17, source rows 38029–62841	2026-10-17, source rows 62892–86343
raw → kept	500 → 434	500 → 417 (23 cross-batch pixel collisions excluded, 8 within-batch duplicates, 42 e-mail PII pages removed, 10 pending-review pages removed)	500 → 433 (29 cross-batch pixel collisions excluded, 6 within-batch duplicates, 26 e-mail PII pages removed, 6 pending-review pages removed)	500 → 430 (25 cross-batch pixel collisions excluded, 8 within-batch duplicates, 25 e-mail PII pages removed, 12 pending-review pages removed)
split	405 train / 29 development, 125 documents	357 train / 60 development, 140 documents	401 train / 32 development, 117 documents	346 train / 84 development, 109 documents
label review	6 needs-human-review (image-checked)	none — 10 flagged rows (3 pixel conflicts, 7 short labels) were removed, not corrected; ledger removed-needs-review.jsonl	none — 6 flagged rows (5 pixel conflicts, 1 short label) were removed, not corrected; ledger removed-needs-review.jsonl	none — 12 flagged rows (8 pixel conflicts, 4 short labels) were removed, not corrected; ledger removed-needs-review.jsonl

Documents that appear in more than one batch (7 between batches 1–2, 6 between batch 3 and its predecessors, 3 between batch 4 and its predecessors) are always assigned the same split side by the deterministic document-hash rule, so no document straddles train/development. Cross-batch exclusion covers exact RGB-pixel identity only; near-duplicate screening between batches was not performed. Combined overlap screening against the seven local eval/train manifests: 0 hits for all four batches.

Cleaning performed (batch 1, 2026-10-14)

• Exact-pixel deduplication: 8 duplicates dropped (7 cross-document, 1 within one document); ledger in dropped-duplicates.jsonl.
• 3 same-pixel label conflicts resolved by visual check: the more complete label kept, rows flagged needs-human-review.
• 1 image-verified correction (interleaved OCR garble in a table cell, source row 5458); before/after hashes in corrections.json.
• 4 visually confirmed incomplete labels flagged needs-human-review (table bodies not transcribed); not rewritten.
• Overlap screening vs 7 local eval/train manifests (IMPACT history test A, full-page dev v4, history line corpus v2, EHRI printed replay, synthetic train A/B/C): 0 exact-image, exact-text, near-image (dHash ≤ 4) or near-text (7-word shingle ≥ 0.9) hits. See overlap-report.json.
• Document-level split (125 documents, disjoint): 405 train / 29 development rows.
• PII removal (2026-10-17): 58 pages whose labels contained e-mail addresses were removed (removed-pii.jsonl, matches masked); checksummed PESEL, mod-97 PL accounts and +48 phones matched nothing.

Boundaries — read before use

• Labels are automated extractions. Nothing here is human-verified or gold. 428 rows are unreviewed-ai, 6 are needs-human-review.
• training_eligible and evaluation_eligible are false for every row by design.
• Overlap screening is bounded to the listed local manifests; upstream pretraining overlap is unaudited.
• Synthetic manifests were screened on exact hashes only.
• PII removal used deterministic high-precision patterns on the labels only: names, street addresses and identifiers without checksums remain, and PII visible only in page images is not detected. Treat the set as containing possible personal data.

License and attribution

• The upstream dataset ahmedheakl/docatlas_instruct is declared CC BY 4.0 by its authors at the pinned revision (license: cc-by-4.0 in the dataset card and Hub metadata), with the statement that source documents were collected from publicly accessible web content and academic repositories under permissive licenses. Per-document license evidence is not published upstream; this subset inherits the upstream dataset-level declaration.
• Modifications here: exact-pixel deduplication, one label correction, document-level split and screening metadata. Page images and labels are otherwise unmodified.
• Required attribution when reusing: Ahmed Heakl et al. (2026), DocAtlas: Multilingual Document Understanding Across 80+ Languages, arXiv:2605.12623; dataset ahmedheakl/docatlas_instruct (CC BY 4.0).

Files

• original-images/ — 434 PNG page images (byte-identical to the source snapshot; image_sha256 binds each manifest row).
• manifest.jsonl — one row per kept page: ids, provenance, text, hashes, split, label status, eligibility flags.
• dropped-duplicates.jsonl — removed duplicates with their kept counterpart.
• removed-pii.jsonl — pages removed for e-mail addresses in labels (matches masked).
• corrections.json, overlap-report.json, report.json, checksums.json — evidence.
• prepare_clean_corpus.py, remove_pii.py — deterministic pipeline scripts (batch 1).
• batches/batch2-20261017/, batches/batch3-20261017/, batches/batch4-20261017/ — batch packages: images, manifest, ledgers (cross-batch exclusions, duplicates, PII, removed pending-review pages), overlap report, report, checksums, pipeline script.