Datasetpublic

PiotrSty/ehri-pl-lines

ehri-pl-lines Line-level OCR/HTR dataset of Polish typewritten historical documents, derived from the Polish sub-corpus of the EHRI dataset. Each example is a single text-line crop plus its ground-truth transcription. Line bounding boxes come from the original ALTO XML ground truth (no automatic detection was used), so transcriptions are reliable and aligned. Content and provenance 468 line crops from 15 pages across 6 documents (ZIH collection). Source images… See the full description on the dataset page: https://huggingface.co/datasets/PiotrSty/ehri-pl-lines.

sourceHugging Facecc-by-4.0updated 8d agoView on Hugging Face

0likes371downloads

Dataset Card

ehri-pl-lines

Line-level OCR/HTR dataset of Polish typewritten historical documents, derived from the Polish sub-corpus of the EHRI dataset.

Each example is a single text-line crop plus its ground-truth transcription. Line bounding boxes come from the original ALTO XML ground truth (no automatic detection was used), so transcriptions are reliable and aligned.

Content and provenance

—468 line crops from 15 pages across 6 documents (ZIH collection).
—Source images and ALTO ground truth: EHRI project, held at the Archives of the Jewish Historical Institute (ŻIH), Warsaw.
—These are early Holocaust testimonies. The material is historically sensitive; it is redistributed here only for OCR/HTR research, consistent with the original CC-BY 4.0 license and the EHRI dataset's intended use.

Split (by document, no leakage)

Splits are assigned by whole document so that test documents are never seen in training.

Split	Documents	Lines
train	ZIH3010106, ZIH3010884, ZIH3010945	349
dev	ZIH3010905	38
test	ZIH3010965, ZIH3010201	81

Files

—train/, dev/, test/ — <id>.png line crop + <id>.txt UTF-8 transcription.
—manifest.jsonl — one record per line: id, split, document, source_page, image, text_file, text, source_bbox ([x1,y1,x2,y2] in source pixels), sha256 of the crop.
—ehri-pl-lines-v1.tar.gz — single-archive copy of all splits + manifest, to avoid per-file API rate limits when downloading on training platforms.

Normalization

Transcriptions are Unicode NFC with whitespace collapsed; case and diacritics are preserved.

License and attribution

Released under CC-BY 4.0, inheriting the license of the source EHRI dataset. Please cite the EHRI project and the Jewish Historical Institute (ŻIH) archives when using this data.