polocrbench-train-pool-v1

A single self-contained training pool for Polish document understanding, assembled for work toward PolOCRBench (PolEval 2026 — transcription, table extraction, KIE). Community-assembled training material — not the official organizer release.

3 551 pages (3 346 train / 205 development) from five open-licensed sources:

source	pages	license	labels
docatlas-pl-clean (repo)	1 713	CC BY 4.0 (Heakl et al. 2026, arXiv:2605.12623)	full-page Markdown
impact-hist-train (repo)	298	CC BY 3.0 (IMPACT Centre of Competence / PSNC)	page texts joined from 1 987 line transcriptions
polocrbench-synth (project-authored, seed 20260922)	1 200	CC BY 4.0	Markdown + HTML tables + KIE fields (faktura/umowa/pismo_urzedowe/formularz)
polocrbench-synth-photo (project-authored, seed 20261018, photo recipe)	300	CC BY 4.0	same, with phone-photo degradations: perspective warp, directional shadow ramp, blur 0.8–2.0, downscale + JPEG
handwriting-notes (author self-licensed: Piotr Styła, 2026-10-17)	40	CC BY 4.0	full-page transcription incl. bracketed drawing descriptions; 16 pages orientation-normalized (rot. 180°)

Synthetic pages carry the subtask-B and C annotations: tables.jsonl (1 072 HTML tables) and kie.jsonl (1 500 field-level JSON records), keyed by page_id. kie-real.jsonl adds precision-first KIE on 45 real contemporary pages (29 checksum-valid NIP values, 6 labeled dates, 20 labeled amounts; deterministic extraction, ISO dates / dot-decimal amounts, spec-aligned) — the contemporary corpus is genuinely sparse in these identifiers, so this is an honest thin auxiliary layer, not a substitute for the synthetic KIE mass. Historical pages are out of KIE scope (the spec's KIE types are invoice/contract/official letter/form). Together the synthetic sources cover the task's planned training composition (~1 200 synthetic) including the photo-degradation family the spec asks for; real pages (2 011) exceed the planned ~800 — an intentional, documented deviation.

Quality gates

• Cross-union dedup on file-sha256 and text-sha256: 1 dropped (ledger dropped-dedup.jsonl).
• Deterministic PII label scan (e-mail, +48 phone, checksummed PL PESEL, mod-97 PL accounts) on real sources only — synthetic identities are generated data and PESEL/NIP fields are legitimate training signal: 0 removed (dropped-pii.jsonl).
• Overlap screening vs four frozen local eval manifests: 0 against IMPACT Test A staged, full-page dev v4 and EHRI printed replay. The 208 flags against qwen3vl-hist-corpus-v2 hit only its TRAIN pages (same source works; 65 pages intentionally shared via the psnc subset) — zero flags against its validation/test splits. See overlap-report.json.
• Splits are document-consistent: docatlas development rows (205 pages) carried over from the source manifests; synthetic and historical are train-only; frozen eval collections are excluded by construction.

Boundaries

• Nothing is human-verified or gold. Labels are automated (DocAtlas, IMPACT ground truth joined per page), generator-produced, or AI pre-annotations of the author's handwritten notes (drawings described in brackets; the author may verify against his own notes).
• The official PolOCRBench Train/Test A/Test B, evaluator and schemas come from the organizers (poleval/2026-polocrbench); this pool is independent training material.
• Historical page JPEGs are format-converted derivatives of the IMPACT dLibra TIFFs (see the source repo for the acquisition audit).

Files

• images/ — 3 511 page images, named by pool id (da<batch>_<row>, hist_<work>__<doc>, syn_<id>, synph_<id>).
• manifest.jsonl — one row per page: id, source, image/text hashes, split, license, provenance flags.
• tables.jsonl, kie.jsonl — subtask B/C annotations for synthetic pages.
• kie-real.jsonl, kie-real-report.json — deterministic KIE mining on real pages (mine_kie_real.py).
• dropped-*.jsonl, overlap-report.json, report.json, checksums.json — evidence; assemble_train_pool_v1.py — deterministic assembly script.