StatMapCorpus v2

23,549 English-language statistical maps from 16,374 web domains, identified in MapPool and annotated by thematic map type with vision-language models.

This repository is a mirror. The citable source of record is the deposit in Dane Badawcze UW (University of Warsaw): https://doi.org/10.58132/FSJGSP, version 2.0. The thirteen files in deposit/ are byte-identical to that release; deposit/schema.json carries their SHA-256 sums so you can verify this yourself. Cite the DOI, not this URL.

The deposit contains annotations, identifiers, image embeddings and documentation. No images are redistributed: each record points to an image on its original host, and an included script retrieves the images and verifies them against the checksums recorded at annotation time.

• Code: https://github.com/MartaSolarz/StatMapCorpus (MIT), archived at https://doi.org/10.5281/zenodo.23215421.

Before you use it

• The labels are a silver standard. Against two human coders, precision was 0.90–0.97 for choropleths, diagrams and cartograms, 0.63–0.71 for heat maps, dot density and flow maps, and 0.52 for isolines (deposit/VALIDATION.md). The errors are systematic, not random noise. Labels for these four types should be verified before they are used as ground truth; deposit/validation.csv lets you estimate the error rate for a chosen subset.
• Splits. Group records by phash, not by duplicate_group_id: 3,294 rows share a perceptual hash with another row without being byte-identical. Grouping also by domain further reduces leakage between maps from the same publisher.
• Class balance. The type labels are multi-label and highly imbalanced. Choropleth covers 88.7% of records, and each of the five rarest types covers less than 1.5%. Report macro-averaged metrics.
• Legend type. A null value in legend_is_classed means not applicable (no legend, a qualitative legend or an illegible one), not a missing value.
• Composition. The corpus is shaped by the selection design: a web crawl, a classifier threshold of 0.80 and at most five images per domain. It is not a representative sample of statistical maps on the web, and it is not exhaustive even among the candidates examined.

Files

Everything under deposit/ is the published release, unmodified.

File	Contents
statmapcorpus_dataset.parquet	The main table: one row per map, 23,549 rows, 30 columns.
embeddings_clip_vitl14.npz	CLIP ViT-L/14 image embeddings, float32, 23,549 × 768, keyed by uid.
candidates.parquet	All 203,371 candidates sampled in Stage 3, with the outcome of each step from the URL check to Gate 4.
validation.csv	Labels of two human coders for the 157-map validation sample, model labels from two repeated runs of Gates 3 and 4 for 300 maps, and a blind audit of 150 images.
positive_keywords.json, negative_keywords.json	The Stage 1 dictionaries: phrases and their weights.
download_images.py	Retrieves the images and verifies them by SHA-256 and perceptual hash.
schema.json	Field names, data types and SHA-256 checksums of all files.
CODEBOOK.md	Every column of every data file: type, permitted and observed values, caveats.
VALIDATION.md	Results of the technical validation.
DATASHEET.md	The dataset documented following Datasheets for Datasets.
NOTICE.md	Provenance, licences and the procedure for removal requests.
README.md	The deposit's own README, as published.

The dataset viewer shows the main table only. The other files have to be loaded directly.

from huggingface_hub import hf_hub_download
import numpy as np, pandas as pd

repo = "msolarz/StatMapCorpus"
df = pd.read_parquet(hf_hub_download(repo, "deposit/statmapcorpus_dataset.parquet",
                                     repo_type="dataset"),
                     dtype_backend="numpy_nullable")
npz = np.load(hf_hub_download(repo, "deposit/embeddings_clip_vitl14.npz",
                              repo_type="dataset"))
uids, emb = npz["uids"], npz["emb"]          # emb.shape == (23549, 768)

How the corpus was built

MapPool (Schnürer 2024) holds 75.9 million map-like images from the CommonPool collection of DataComp, each with its URL, alt-text and CLIP embedding. StatMapCorpus was selected from it in four stages, ordered by increasing cost per record:

1. Keyword filter. Weighted phrase dictionaries in 21 languages are matched against the alt-text; records scoring at least −1 are retained (54.9 million).
2. Image classifier. An SVM with a radial basis function kernel on the CLIP ViT-L/14 embeddings, trained with active learning; records with a predicted probability of at least 0.5 are passed on (about 1.68 million).
3. Deduplication, sampling and retrieval. URLs are normalised and deduplicated; records with a probability of at least 0.80 are sampled, at most five per domain (203,371 candidates); the URLs are checked and the images downloaded (116,573 images).
4. Vision-language cascade. Gate 1 (rule): the shorter side is at least 400 px. Gate 2 (Claude Haiku 4.5): a map covering at least about half of the image. Gate 3 (Claude Haiku 4.5): a statistical map with quantitative data referenced to administrative or statistical units, with English lettering. Gate 4 (Claude Sonnet 4.6): at least one of the seven thematic map types, the legend type and the model's confidence.

The images were downloaded on 1 August 2026 and annotated on 1–2 August 2026. Each map carries seven non-exclusive thematic map type labels (choropleth, diagram, isoline, dot density, heat map, cartogram, flow map), a legend type, a one-sentence description, the model's self-reported confidence, image properties, SHA-256 and perceptual hashes, and a CLIP image embedding. The alt-text is not redistributed; it belongs to MapPool and can be joined on uid.

Getting the images

cd deposit
python download_images.py --data statmapcorpus_dataset.parquet --out images/ --verify-phash

Every downloaded file is checked against the sha256 recorded at annotation time. Where the hash differs but the 64-bit phash still matches, the host has only re-encoded the file and the labels still apply. Where neither matches, the image has changed and its labels should not be used. All 23,549 images were retrievable on 1 August 2026, and the share of reachable URLs will decline over time.

Licence

Annotations, metadata and documentation are CC BY 4.0. The CLIP embeddings are derived from MapPool and are also CC BY 4.0. download_images.py is MIT; the Hub records a single licence per repository, so the MIT grant is stated here and in the file's own header. Rights in the images remain with their owners; they are neither redistributed nor licensed here.

deposit/NOTICE.md gives the full attribution, the derivation chain from MapPool and CommonPool/DataComp, and the procedure for removal requests. Removal requests are handled against the deposit, and this mirror is updated to match.

Citation

@dataset{solarz2026statmapcorpus,
  author    = {Solarz, Marta},
  title     = {{StatMapCorpus}: 23,549 English-language statistical maps from {MapPool},
               annotated by thematic map type using vision-language models},
  year      = {2026},
  publisher = {Dane Badawcze UW},
  version   = {V2},
  doi       = {10.58132/FSJGSP},
  url       = {https://doi.org/10.58132/FSJGSP}
}

Please also cite MapPool, from which the identifiers and embeddings are inherited: Schnürer, R. (2024). MapPool – Bubbling up an extremely large corpus of maps for AI. Abstracts of the ICA 8, 22. https://doi.org/10.5194/ica-abs-8-22-2024

Versions

Tag	Deposit version	Notes
v1.0	V1 (2026-08-24)	Initial release. Model-generated annotations, no human validation.
v2.0	V2	Validation files, candidate list, datasheet; "cartographic method" renamed to "thematic map type"; file names without a version number. The main table has the same columns and values as in V1.

Contact

Marta Solarz, Faculty of Geography and Regional Studies, University of Warsaw – m.solarz2@uw.edu.pl

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.