StatMapCorpus contains 23,549 English-language statistical maps identified in MapPool (Schnürer 2024), a collection of 75.9 million map-like images drawn from CommonPool/DataComp. Candidates were narrowed in four stages: a multilingual keyword filter over the alt-text inherited from MapPool (score ≥ −1.0), an SVM-RBF classifier over CLIP ViT-L/14 embeddings trained with uncertainty sampling, URL validation and image retrieval for candidates sampled at pred_proba ≥ 0.80 (at most five per domain), and a cascade of vision-language model gates. An image was included if the map occupies at least about half of the image, represents quantitative data referenced to administrative or statistical units, is lettered in English, and uses at least one of seven thematic map types. It also had to be retrievable and at least 400 px on its shorter side.
The deposit contains annotations, identifiers, image embeddings and documentation; no images are redistributed. Each record carries the image URL and file-verification surrogates (sha256, 64-bit pHash) computed from the exact file that was annotated, and an included script retrieves the images and verifies them against the published checksums. All records were retrieved on 1 August 2026 and annotated on 1–2 August 2026; because the population comes from a web crawl, the reachable fraction can decrease from that date. Alongside the corpus, candidates.parquet documents the 203,371 candidates that entered the retrieval stage and the outcome of every step applied to them, and validation.csv holds the human codings, repeated model runs and audit answers behind the validation summarised below.
The annotation layer provides a multi-label thematic map type taxonomy (choropleth, diagram, isoline, dot density, heat map, cartogram, flow map), a legend-type field (classed, continuous or not applicable), a short model-generated content description, and a CLIP ViT-L/14 embedding inherited from MapPool. Type labels are strongly skewed: choropleth covers 88.7% of records and five of the seven types fall below 1.5%. Byte-identical copies are flagged; near-duplicates can be recovered from the published pHash, and CODEBOOK.md documents every column, defines each duplication measure, and explains how to group records when making splits.
Labels are model-generated and were validated on samples rather than exhaustively. Two coders, working independently in an interface that showed no model output and following a written codebook, coded 157 maps across 12 questions; they disagreed in 99 of 1,874 paired judgments. Precision of the type labels was high for choropleth, diagram and cartogram (0.90–0.97), lower for heat map, dot density and flow map (0.63–0.71) and lowest for isoline (0.52); the complete set of seven type labels matched the coders' set for 0.94 of maps. Re-running Gates 3 and 4 twice for 300 maps left the complete type set unchanged in all three runs for 99.3% of a random subsample, so the errors in the rare types are systematic rather than noise and would not be removed by re-running. A blind audit of 130 rejected images, with 20 corpus maps as controls, indicates that roughly 2,600 images meeting the criteria were lost in the cascade, almost all at the gate judging whether a map is statistical. VALIDATION.md gives the full results.
Three limitations follow. The corpus is not a random sample of statistical maps: it inherits the composition of a web crawl and the selection criteria above, and it is not exhaustive even among the candidates examined. Analyses of the rare types should allow for their lower precision. The 17 maps carrying no legible lettering are in the corpus incidentally, because the language gate required English lettering, and they do not represent textless maps systematically.
Annotations, metadata and documentation are licensed CC BY 4.0; the CLIP embeddings are CC BY 4.0 and require attribution to MapPool; the retrieval script is MIT. Rights in the images remain with their owners and the images are neither redistributed nor licensed here. NOTICE.md sets out attribution, the derivation chain and the takedown procedure.
(2026-10-07)
