---
pretty_name: Geospatial data use in World Bank PADs and PIDs
configs:
- config_name: projects
  data_files: data/projects.jsonl
- config_name: documents
  data_files: data/documents.jsonl
- config_name: passages
  data_files:
  - {split: train, path: data/passages/train.jsonl}
- config_name: related_mentions
  data_files:
  - {split: train, path: data/related_mentions/train.jsonl}
---

# Geospatial data use in World Bank PADs and PIDs

The data behind [How often do World Bank projects use geospatial data?](https://rafmacalaba.github.io/pad-pid-geospatial-analysis.html).
Snapshot **`938866a81eab`**, annotations to 2026-09-25 21:57 UTC. The page shows the same snapshot id.

## Headline

| | Out of: those that use data | Out of: all |
|---|---|---|
| Projects | 1,165 of 1,932 (60.3%) | 1,165 of 2,261 (51.5%) |
| Documents | 1,637 of 3,544 (46.2%) | 1,637 of 4,494 (36.4%) |

A project uses geospatial data when at least one data mention in the reviewed passages of any of its PADs/PIDs
is judged *used* (`annotations[].data_use.used`) and carries a place label (`geospatial_class` is
`point_geocoded`, `admin_linked` or `national_spatial_derived`). Repeat copies of a document share a
`distinct_document_id`, and a passage read twice counts once. `projects.jsonl` has the result per project.

## Scope

Review sets umar_pads, umar_pads_500; document collection umar_pads: 4,708
PADs/PIDs from 2,261 projects, every one with at least one reviewed passage. Each document was checked
against the World Bank documents catalogue (`catalogue_check`: exists, has the stated type, belongs to the stated project).

## Files

- `projects`: one row per project, with country, World Bank country code (regional and multi-country projects carry codes such as `3W`), current status, region, primary sector, approval date (World Bank Projects API) and the page's yes/no results.
- `documents`: one row per reviewed document id, with its catalogue record and check.
- `passages`: every reviewed passage with all annotations. `data_use` is the verdict the page counts (owner ruling first, else the agent's proposal); `prior_data_use_rulings` is the snapshot taken when the passage was annotated.
- `related_mentions`: the place-linked mentions only, with passage-aligned `render_*` offsets.
- `metadata/summary.json`: the four headlines; `metadata/manifest.json`: file hashes. `reports/`: the published pages.

## Limits

These are AI-extracted data-use mentions from a sample of passages: an AI model found each data mention, judged
whether the data was used, and labelled its geography against a written rubric, quoting the passage words as evidence.
Only sampled passages of each document were read. See the page's "How reliable are these figures?" note for the
stricter-check range.