deposit/README.md

# 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.

- **Dataset (version of record):** https://doi.org/10.58132/FSJGSP – Dane Badawcze UW,
  version 2.0. Version 1 remains available under the same DOI.
- **Data descriptor:** Solarz, M. *A corpus of English-language statistical maps from the
  web annotated by thematic map type.* Scientific Data (submitted).
- **Code:** https://github.com/MartaSolarz/StatMapCorpus (MIT): the pipeline, the prompts,
  the export of these files and the analyses of the validation.
- **Mirror of the main table:** https://huggingface.co/datasets/msolarz/StatMapCorpus

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.

## Files

| File | Format | Contents |
|---|---|---|
| `statmapcorpus_dataset.parquet` | Apache Parquet | The main table: one row per map, 23,549 rows, 30 columns. Identifiers and provenance, image properties and checksums, labels. |
| `embeddings_clip_vitl14.npz` | NumPy NPZ | CLIP ViT-L/14 image embeddings, float32, 23,549 × 768, keyed by `uid`. |
| `candidates.parquet` | Apache Parquet | All 203,371 candidates sampled in Stage 3, with the outcome of each step from the URL check to Gate 4. |
| `validation.csv` | 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` | JSON | The Stage 1 dictionaries: phrases and their weights. |
| `download_images.py` | Python | Retrieves the images from their URLs and verifies them by SHA-256 and perceptual hash. |
| `schema.json` | JSON | Field names, data types and SHA-256 checksums of all files. |
| `CODEBOOK.md` | Markdown | Every column of every data file: type, permitted and observed values, caveats. |
| `VALIDATION.md` | Markdown | Results of the technical validation. |
| `DATASHEET.md` | Markdown | The dataset documented following *Datasheets for Datasets*. |
| `NOTICE.md` | Markdown | Provenance, licences and the procedure for removal requests. |

## What changed from version 1

- **New files:** `candidates.parquet`, `validation.csv`, `VALIDATION.md` and
  `DATASHEET.md`.
- **The labels have been validated.** They were checked against two human coders, by
  repeating the model calls and by an audit of rejected images (`VALIDATION.md`).
  Version 1 stated that no human validation had been done.
- **File names no longer carry a version number.** The version is that of the deposit
  (2.0). `statmapcorpus_dataset_v1.parquet` of version 1 is now
  `statmapcorpus_dataset.parquet`.
- **The main table is unchanged.** It has the same columns and values as in version 1;
  only the file name and the `dataset` entry of its metadata changed. The embeddings file
  and the dictionaries are unchanged.
- **Terminology follows the data descriptor.** "Cartographic method" is now "thematic map
  type". The column names `method_*` and `methods_count` are kept, so that code written
  for version 1 still runs. The codebook gives the definition of each type (Table 3 of the
  data descriptor) and the numbering of the gates.

## 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): the image shows a map that covers at least about half
     of it.
   - **Gate 3** (Claude Haiku 4.5): a statistical map with quantitative data referenced to
     administrative or statistical units, with English lettering. Phenomena measured by
     sensors or physical models — weather, climate, hydrology, pollution, geology,
     ecology — are excluded.
   - **Gate 4** (Claude Sonnet 4.6): at least one of the seven thematic map types; it also
     records the legend type and its confidence.

The images were downloaded on 1 August 2026 and annotated on 1–2 August 2026.
`candidates.parquet` records the outcome of every step of Stages 3 and 4.

Each map carries seven non-exclusive thematic map type labels: choropleth, diagram,
isoline, dot density, heat map, cartogram and flow map. It also has a legend type
(classed, continuous or not applicable), 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

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

Every downloaded file is checked against the `sha256` recorded at annotation time:

- **The hash matches:** the file is the one that was annotated.
- **The hash differs but the 64-bit `phash` matches:** the host has only re-encoded the
  file, and the labels still apply.
- **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. The embeddings remain available for every record, but they were
computed on images downscaled to at most 512 px.

## 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 (`VALIDATION.md`). The errors are systematic, not random
  noise. Labels for these four types should be verified before they are used as ground
  truth; `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. Do not treat it as a
  continuous legend.
- **Maps without lettering.** The 17 maps without lettering are in the corpus only
  incidentally, because Gate 3 required English lettering. They are not a sample of
  textless maps.
- **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: an audit suggests that about 2,600 qualifying images were rejected, almost all
  at Gate 3.

## Licence, citation and contact

Annotations, metadata and documentation are licensed CC BY 4.0. The CLIP embeddings are
derived from MapPool and are also licensed CC BY 4.0. `download_images.py` is MIT. Rights
in the images remain with their owners. `NOTICE.md` gives the full attribution, the
derivation chain and the procedure for removal requests.

Please cite the dataset and the data descriptor:

> Solarz, Marta, 2026, "StatMapCorpus: 23,549 English-language statistical maps from
> MapPool, annotated by thematic map type using vision-language models",
> https://doi.org/10.58132/FSJGSP, Dane Badawcze UW, V2.

Marta Solarz, Faculty of Geography and Regional Studies, University of Warsaw –
m.solarz2@uw.edu.pl · ORCID 0009-0001-4134-8500

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.