HaLO: Handwriting Assessment for Legibility and Ordering

HaLO is a dataset of 1,907 handwriting images from 202 German schoolchildren (ages 10--12), annotated with 19,070 pairwise legibility judgments. It supports research on AI-based handwriting legibility assessment through comparative ranking.

Dataset Description

Each handwriting sample depicts one of ten predefined German sentences (20--27 characters each), written by a child in their natural handwriting style. The samples were scanned and cropped from paper sheets. Due to privacy constraints, only Pixtral 12B image embeddings (1024-dimensional feature vectors) are released, not the raw images.

Annotation

Legibility annotations were collected using a pairwise comparison paradigm: annotators were shown two handwriting samples side by side and asked to select the more legible one. A pre-study compared this approach with absolute Likert-scale ratings and found that pairwise comparisons produce higher inter- and intra-rater agreement. The main study annotations were collected by eight annotators (three experts, five non-experts).

Data Splits

The dataset is split into training (1,134 samples, 121 children), validation (382 samples, 40 children), and test (391 samples, 41 children) partitions. All samples from the same child are assigned to the same partition to prevent information leakage.

Configurations

default / main_study

Pairwise legibility annotations from the main study. Each row is a comparison between two samples.

Field	Description
sampleId1, sampleId2	IDs of the two compared samples
score	1 if sample 1 is more legible, -1 if sample 2 is more legible
samplePath1, samplePath2	Paths to the corresponding feature .npy files
userId	Annotator ID
referenceSentenceId1, referenceSentenceId2	Reference sentence IDs (1--10)
questionId	Annotation question ID
submissionTimestamp	Timestamp of the annotation

Splits: train (11,340), test (3,910), validation (3,820)

from datasets import load_dataset
dataset = load_dataset("MarcoLents/HaLO")

pre_study

Pairwise legibility annotations from the annotation pre-study, with the same schema as the main study. Splits correspond to individual annotators, where Annotator 1 performed the procedure twice (splits annotator11 and annotator12).

Splits: annotator11 (1,839), annotator12 (1,530), annotator2 (1,606), annotator3 (1,541)

pre_study = load_dataset("MarcoLents/HaLO", "pre_study")

characteristics

Binary sample characteristics annotated per sample.

Field	Description
sampleId	Sample ID
text	Transcribed text content
referenceSentenceId	Reference sentence ID
isWrittenInPureCursive	Whether the sample is written entirely in cursive
isStrokeThin	Whether the stroke width is predominantly thin
containsTypo	Whether at least one typographical error is present
containsCorrection	Whether at least one correction is present
samplePath	Path to the feature .npy file

Splits: train (1,134), test (391), validation (382)

characteristics = load_dataset("MarcoLents/HaLO", "characteristics")

absolute_pre_study

Absolute legibility ratings on a 5-point Likert scale (0 = very legible, 4 = not legible at all) from the annotation pre-study. Four annotators independently rated a random subset of 136 samples. Annotator 4 rated twice (annotator41, annotator42).

Field	Description
sampleId	Sample ID
referenceSentenceId	Reference sentence ID
annotator41 -- annotator7	Integer Likert ratings (0--4) per annotator
samplePath	Path to the feature .npy file

Splits: train (136)

absolute_pre_study = load_dataset("MarcoLents/HaLO", "absolute_pre_study")

absolute_test

Absolute legibility ratings on a 5-point Likert scale (0 = very legible, 4 = not legible at all) from the test set. Four annotators independently rated samples from the test partition. Not all annotators rated every sample; missing values indicate unrated samples. Where a rater provided multiple ratings for the same sample, the value is the rounded mean.

Note: The absolute_pre_study and absolute_test configs use different annotator numbering schemes because they originate from separate annotation campaigns described in different sections of the paper. Some of the underlying raters overlap between the two configs, but the pseudonyms are independent.

Field	Description
sampleId	Sample ID
referenceSentenceId	Reference sentence ID
annotator8 -- annotator11	Integer Likert ratings (0--4) per annotator, NaN if not rated
samplePath	Path to the feature .npy file

Splits: train (391)

absolute_test = load_dataset("MarcoLents/HaLO", "absolute_test")

aspect_ratio

Image aspect ratio for each sample.

Field	Description
sampleId	Sample ID
aspectRatio	Width-to-height ratio of the original image
samplePath	Path to the feature .npy file

Splits: train (1,134), test (391), validation (382)

aspect_ratio = load_dataset("MarcoLents/HaLO", "aspect_ratio")

Feature Files

Pixtral 12B image embeddings are stored as .npy files under features/{referenceSentenceId}/{sampleId}.npy. Each file contains a 1024-dimensional float32 vector obtained by mean-pooling the token-level outputs of the frozen Pixtral-ViT encoder.

To download the feature files:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="MarcoLents/HaLO",
    repo_type="dataset",
    allow_patterns="features/*/*.npy",
    local_dir="data"
)

Citation

If you use this dataset, please cite:

@article{bauer2025halo,
  title={Ranking handwriting images like a human: AI-based legibility assessment by comparative ranking on the HaLO dataset},
  author={Bauer, Meike and Lents, Marco and Schmidt, Erik and Hamann, Tim and Pieger, Lukas and Salata, Susanne and Di Salvo, Francesco and Hoffmann, Tal and Barth, Jens and Ledig, Christian},
  year={2025}
}

License

This dataset is released under the MIT License.

Acknowledgements

This study was supported by the Hightech Agenda Bayern (HTA) of the Free State of Bavaria, Germany. Additional support was provided by the Bavarian State Ministry of Economic Affairs, Regional Development and Energy (StMWi) as part of the Bavarian Collaborative Research Programme (BayVFP), funding line Digitalization, project KIBEL, grant number DIK0813.