deposit/DATASHEET.md

# DATASHEET

StatMapCorpus documented following *Datasheets for Datasets* (Gebru et al. 2021,
https://doi.org/10.1145/3458723). The answers are short. The data descriptor, `README.md`,
`CODEBOOK.md` and `VALIDATION.md` give the details.

## Motivation

**For what purpose was the dataset created?** Vision-language models are increasingly
expected to read statistical maps. Research on map reading, map-type classification and
cartographic quality needs maps that circulate in practice, in sufficient numbers and with
consistent annotations. Existing map datasets are small, synthetically rendered, or not
labelled by thematic map type. StatMapCorpus fills that gap with real web maps labelled
with the seven thematic map types of cartographic textbooks.

**Who created it, and who funded it?** Marta Solarz, Faculty of Geography and Regional
Studies, University of Warsaw. The work was funded by the University of Warsaw under the
Excellence Initiative – Research University (IDUB) programme, through a microgrant for
doctoral students (Action IV.4.1), BOB-IDUB-622-272-2026.

## Composition

**What do the instances represent?** Each instance is one statistical map published on the
web: a map of quantitative data referenced to administrative or statistical units. The
instance is the record of the image (URL, checksums, properties) and its labels; the image
itself is not included.

**How many instances are there?** 23,549 maps from 16,374 web domains.
`candidates.parquet` additionally lists the 203,371 candidates examined, with the
outcome of each selection step.

**Is the dataset a sample?** Yes, and not a representative one. It is a selection from
MapPool, itself a classifier-based selection from a web crawl, and it is shaped by the
selection design:
- a classifier probability of at least 0.80;
- at most five images per domain;
- English lettering;
- a shorter side of at least 400 px.

It is not exhaustive even among the candidates examined: an audit estimates that about
2,600 qualifying images were rejected (`VALIDATION.md`).

**What data does each instance consist of?** The main table holds:
- identifiers and provenance: MapPool `uid`, URL, domain, Stage 1 and 2 scores;
- image properties: dimensions, format, size, SHA-256 and perceptual hash, retrieval
  time;
- labels: a one-sentence description, whether the map carries lettering, seven
  non-exclusive thematic map type labels, the legend type and the model's confidence;
- duplicate flags.

Each instance also has a CLIP ViT-L/14 image embedding. `CODEBOOK.md` defines every field.

**Is there a label or target?** Yes: the seven thematic map type labels (choropleth,
diagram, isoline, dot density, heat map, cartogram, flow map) and the legend type
(classed, continuous or not applicable). They were assigned by a vision-language model
and are a silver standard.

**Is any information missing?** The legend type is null for 6,053 maps. Null means *not
applicable*, not missing. The images themselves are not distributed, and the share of
images still reachable at their URLs will decline over time.

**Are there recommended data splits?** No fixed splits. Group records by `phash` (and
preferably by `domain`) when splitting, so that re-encoded copies of the same map do not
fall on both sides (`CODEBOOK.md`).

**Are there errors, sources of noise or redundancies?**
- **Errors.** The labels contain errors, which are measured in `VALIDATION.md`. Precision
  is 0.90–0.97 for choropleths, diagrams and cartograms and 0.52–0.71 for isolines, dot
  density, heat maps and flow maps, with systematic confusions between related types.
- **Redundancies.** 1,050 records are byte-identical copies and are flagged
  (`is_duplicate_byte`). Further near-duplicates can be found through `phash`.

**Is the dataset self-contained?** No. The images are linked, not included, and the
alt-text belongs to MapPool and can be joined on `uid`.

**Does it contain confidential data, or data about identifiable people?** The dataset
consists of maps of aggregated statistics and contains no images. A linked image could
still contain personal data, because CommonPool was collected from the public web.
`NOTICE.md` gives the removal procedure.

## Collection process

**How was the data acquired?** From MapPool (75,890,407 records, each with the image URL,
alt-text and a CLIP embedding computed by DataComp), in four stages of increasing cost:
1. a keyword filter over alt-texts in 21 languages;
2. an SVM image classifier on the embeddings, trained with active learning;
3. URL deduplication, sampling, URL validation and image download;
4. a cascade of four gates: a size rule, then three calls to Claude Haiku 4.5 and Claude
   Sonnet 4.6.

**Over what timeframe?** The URLs were checked on 30 July and 1 August 2026, the images
downloaded on 1 August 2026, and the annotation ran on 1–2 August 2026. The underlying web
crawl (CommonPool) is older.

**Were any ethical review processes conducted?** No. The dataset contains no images and
no data collected from people. The human coding of the validation was done by the author
and one colleague.

## Preprocessing, cleaning and labelling

**Was any preprocessing or labelling done?**
- URLs were normalised and deduplicated, and three domains were excluded.
- Byte-identical images were flagged rather than removed.
- Labels were assigned by vision-language models at Gates 2–4, from prompts published in
  the code repository.

The raw model responses and local file paths are not released.

**Is the software used available?** Yes: https://github.com/MartaSolarz/StatMapCorpus
(MIT), including the prompts and the export and validation scripts.

## Uses

**What is the dataset intended for?**
- training and evaluating vision-language models on map reading and on recognising
  thematic map types;
- sampling real-world maps for user studies and research on cartographic quality;
- studying how statistical data are mapped on the web.

The published candidate list lets the selection be re-run with other conditions, for
example for another language.

**Is there anything users should know to avoid harm or misuse?**
- Treat the labels as a silver standard, and verify labels of the four weaker types
  before using them as ground truth.
- Do not treat the corpus as a representative sample of statistical maps or of the web.
- Do not treat the 17 maps without lettering as a sample of textless maps.
- Use the images according to their owners' rights.

**Are there tasks for which the dataset should not be used?** It should not be used to
estimate the prevalence of thematic map types among statistical maps in general, because
choropleths are over-represented by the selection design.

## Distribution

**How is the dataset distributed?** From Dane Badawcze UW at
https://doi.org/10.58132/FSJGSP (version of record). The main table is mirrored on Hugging
Face at https://huggingface.co/datasets/msolarz/StatMapCorpus.

**Under what licence?**
- annotations, metadata and documentation: CC BY 4.0;
- the embeddings: CC BY 4.0, derived from MapPool;
- the retrieval script: MIT.

Rights in the images remain with their owners (`NOTICE.md`).

## Maintenance

**Who maintains the dataset, and how can they be contacted?** Marta Solarz,
m.solarz2@uw.edu.pl.

**Will the dataset be updated?** Corrections and removals are published as new versions
under the same DOI, with earlier versions kept. Version 2 added the validation files and
the candidate list, and aligned the terminology with the data descriptor (`README.md`).

**Can others extend it?** Yes. The code and prompts allow the pipeline to be re-run on the
published candidates with other conditions, or on other source collections. Contributions
can be proposed through the code repository.