CoolFace
Datasetpublic

Sigurdur/icelandic-ocr-benchmark

Dataset Card for Icelandic OCR Benchmark Dataset Details Dataset Description Icelandic OCR Benchmark is a ground-truth dataset for evaluating OCR accuracy on Icelandic-language documents. It consists of manually transcribed page images with matching layout annotations (text regions, line polygons, baselines) in both ALTO and PAGE XML. Curated by: Sigurdur Haukur Birgisson Language(s): Icelandic (is) License: CC BY-SA 4.0 Dataset… See the full description on the dataset page: https://huggingface.co/datasets/Sigurdur/icelandic-ocr-benchmark.

sourceHugging Facecc-by-sa-4.0updated 10d agoView on Hugging Face
1likes87downloads
Dataset Card

Dataset Card for Icelandic OCR Benchmark

Dataset Details

Dataset Description

Icelandic OCR Benchmark is a ground-truth dataset for evaluating OCR accuracy on Icelandic-language documents. It consists of manually transcribed page images with matching layout annotations (text regions, line polygons, baselines) in both ALTO and PAGE XML.

  • —Curated by: Sigurdur Haukur Birgisson
  • —Language(s): Icelandic (is)
  • —License: CC BY-SA 4.0

Dataset Sources

The page images were sampled from the Icelandic subset of HuggingFaceFW/finepdfs (document id finepdfs-isl-M). Text regions, lines, and baselines were segmented and manually transcribed using eScriptorium, then exported to this dataset.

Uses

Direct Use

Benchmarking OCR engines against ground-truth Icelandic text, and evaluating line/region segmentation against the included polygons and baselines.

Out-of-Scope Use

This is currently a single 100-page document, not a broad corpus — treat it as a benchmark/spot-check sample rather than a general-purpose Icelandic OCR training set. More documents may be added as additional ground truth is transcribed.

Dataset Structure

Each row is one page:

  • —image: the full page scan
  • —doc_id: identifier of the source document (finepdfs-isl-M)
  • —sequence: the page's order within the document
  • —alto: the page's transcription and layout as ALTO XML (v4)
  • —page: the page's transcription and layout as PAGE XML

To uniquely identify a page, use doc_id together with sequence.

The dataset contains 100 pages and about 3,550 transcribed text lines (~33,000 words).

Dataset Creation

Source Data

Page images sampled from the Icelandic subset of HuggingFaceFW/finepdfs.

Annotations

Text regions and lines were segmented and manually transcribed by hand in eScriptorium. alto and page were generated from that transcription; no character-level OCR confidence data is included since transcription was manual.

Who are the annotators?

Sigurdur Haukur Birgisson.

License

This dataset is licensed under CC BY-SA 4.0.