impresso-project/frakturline-testset

Fraktur/Other Text-Line — Test Set

A balanced, held-out evaluation set of 2 000 scanned text-line images (1 000 per class) for the binary task of distinguishing Fraktur (blackletter / Gothic script) from other script (primarily Antiqua / Latin / Roman).

Developed for the Impresso digital humanities project.

---

Dataset Details

Property	Value
Task	Binary image classification
Classes	fraktur, other
Images per class	1 000
Total images	2 000
Image format	WebP (lossless), grayscale text-line crops
Typical size	variable width × ~60 px height
Languages	German, French, Luxembourgish
Source	Historical Swiss and Luxembourgish newspapers
Sampling	Stratified random sample (seed=42) from the training corpus

---

Data Fields

Field	Type	Description
image	PIL.Image	Grayscale text-line crop
label	ClassLabel	"fraktur" or "other"
file_name	string	Repo-relative path (data/test/{label}/{seq_id}.webp) — sequential ID, carries no provenance
source_tag	string	Original filename stem — partially encodes newspaper/date of origin where known; not systematic and should not be relied upon for analysis

---

Usage

from datasets import load_dataset

ds = load_dataset("impresso-project/frakturline-testset", split="test")
print(ds[0])
# {"image": <PIL.Image>, "label": "fraktur", "file_name": "data/test/fraktur/0001.webp", "source_tag": "..."}

Evaluate a model

from datasets import load_dataset
from PIL import Image
import torch
from torchvision import transforms

ds = load_dataset("impresso-project/frakturline-testset", split="test")

# Load your model here ...

correct = sum(
    1 for item in ds
    if predict(item["image"]) == item["label"]
)
print(f"Accuracy: {correct / len(ds):.3f}")

---

Companion Model

The classifier trained and evaluated against this test set is published at:

→ impresso-project/frakturline-classification-cnn

---

Important Caveats

• source_tag is informational only. It reflects original filenames which are partially human-readable (some encode newspaper ID + date) but the naming is not systematic across the corpus. Do not use it as a structured metadata field.
• Held-out status. These images were moved out of the training corpus before training began (sample_testset.py, seed=42). They should not be used for training.
• Class balance. The test set is artificially balanced (1 000/class). The underlying corpus has a higher proportion of Antiqua text. Reported accuracy on this set does not directly reflect real-world class distributions.
• Script scope. other consists primarily of Antiqua but may include mixed-typeface lines, decorative elements, or non-Latin scripts. It is not a pure Antiqua set.

---

Data provenance and upstream licensing

Part of the source material used to construct this dataset was derived from the National Library of Luxembourg (BnL) OCR ground-truth release bnl-ground-truth-newspapers-before-1878.zip. The BnL historical newspapers portal describes these OCR datasets as Public Domain / CC0 and characterizes the Ground Truth Pack as containing 33,000 transcribed text lines, including 19,000 Antiqua and 14,000 Fraktur lines. See also the dataset overview page: BnL Historical Newspapers.

For background on the BnL OCR workflow and data production context, see:

Schneider, Pit, and Yves Maurer. Rerunning OCR: A Machine Learning Approach to Quality Assessment and Enhancement Prediction. Journal of Data Mining & Digital Humanities, 2022. Article page: https://jdmdh.episciences.org/10239.

The present dataset is not identical to the original BnL release. The upstream material was cleaned by removing lines that did not contain identifiable Fraktur or Antiqua characters and by excluding mixed-font cases, while retaining lines containing digits. Curation was performed iteratively, beginning from the original BnL splits and involving repeated manual inspection of prediction errors across cross-validation splits. For example, lines labeled as Fraktur but consistently predicted as Antiqua, or vice versa, were reviewed manually. In some cases these were acceptable borderline instances, but many revealed inconsistencies with the dataset guidelines, which were intended to preserve only clear Fraktur / non-Fraktur distinctions. Such inconsistent cases were removed.

The test set also includes 84 additional lines from impresso corpus material curated by the dataset creators.

---

Release version: v1.0.0
This release may receive metadata-only updates after publication. The dataset contents associated with this release are frozen under the Git tag v1.0.0. Any modification to the dataset contents, including items or labels, triggers a new release version.

--

License

This dataset contains material from multiple sources with different rights statuses.

• Part of the source material derives from the National Library of Luxembourg (BnL) OCR ground-truth release bnl-ground-truth-newspapers-before-1878.zip, which the BnL describes as CC0 / Public Domain.
• Additional original contributions made by the dataset creators in this release — including curation, selection, cleaning decisions, split construction, and documentation — are made available under CC BY 4.0.

Users should therefore distinguish between upstream source material, which remains available under its original terms, and original contributions in this release, which are provided under CC BY 4.0.

This dataset is intended primarily for evaluation use. Do not train models with it! Users remain responsible for assessing the rights status of any upstream source material, especially where historical newspaper provenance or copyright status is not fully documented across jurisdictions.

If you use this dataset, please cite the Impresso project and link to this repository.