polocrbench-official-splits (PRIVATE)

Two-split package of frozen evaluation material offered to the PolOCRBench organizers (poleval/2026-polocrbench). Private by design: evaluation material must not be public while participants test on it (organizer instruction, 2026-10-17). Training material stays public — see the train pointers below.

Split assignment is a proposal by the data contributor; the organizers decide the final splits.

Structure (2 splits)

split	contents	pages
train/	pointer manifest to the public pool PiotrSty/polocrbench-train-pool-v1 (ids + sha256; no images copied — training data is public by design)	3 551
testA/	development-style pages: fullpage dev v4 (15, human-checked) + hist corpus v2 validation (15)	30
testB/	holdout pages: IMPACT print test (36) + hist corpus v2 test (9); plus testB/ehri-lines/ EHRI printed replay line crops	45 pages + 192 lines

Every image was checksum-verified against its source manifest during packaging (PACKAGE-REPORT.json, evidence/).

Provenance and licenses

• Historical pages: IMPACT/PSNC ground truth, CC BY 3.0 (attribution: IMPACT Centre of Competence / PSNC). Sources: PiotrSty/impact-print-v2 (test), PiotrSty/impact-psnc-polish-ocr (subset test split), qwen3vl-hist-corpus-v2 (validation/test splits, derived from the above).
• Fullpage dev v4: single-review reference with scoped user confirmations (see its experiment documentation in PiotrStyla/OCR_engine).
• EHRI printed replay lines: derived from EHRI digitized books; confirm the upstream license statement before any public release of this material.
• Train pointers resolve to a pool licensed CC BY 4.0 / 3.0 (per source).

What the organizers should decide

1. Final split naming and sizes (Test A / Test B semantics per the task card).
2. Whether the EHRI line family belongs in the benchmark at all (line-level, not page-level).
3. Whether dev v4 (human-checked) is good enough to be a scored reference or should stay development-only.