Data

Every published series, as CSV and Parquet, one at a time or all in one file, with the data dictionary and the provenance record beside them. Nothing on this page is typed by hand: the counts and the bundle's byte size are read from the package's own manifest when the page is rendered.

Download everything

Download the whole dataset (.zip) Download the data dictionary (.csv)

15,625,125 bytes (erwin_data_v1.zip, 654 members). It unpacks to one dated folder holding csv/ (324 files) and parquet/ (324 files), plus the citation file and llms.txt; the data dictionary, the provenance record and the terms of use sit at the zip root, and the series registry is under metadata/. Per-member byte counts and SHA-256 hashes are published in BUNDLE_MANIFEST.csv, so the size above can be checked against the file rather than taken on trust.

Formats

CSV and Parquet. JSON is not offered as a data format — the series registry is reachable as metadata at /metadata/series_registry.json, which is a description of the data, not the data.

• data_dictionary.csv — 325 rows, one per published series: name, category, units, frequency, coverage and source.
• PROVENANCE.csv — one row per series: who published it, how it was constructed, how many rows the file really holds, and what shape it is.
• TERMS.md — terms of use. Each upstream publisher keeps its own; SIPRI's in particular restrict commercial redistribution.
• CITATION.cff — how to cite this dataset.
• WEB_MANIFEST.json — metadata: the counts this page renders, so every number here is checkable.

What the package contains

These are separate measures and they are deliberately not all equal. Forcing them to agree would hide exactly what a user needs to know.

Every series

All 325 published series, each with its own downloads. “Shape” is what the file actually is, determined by reading it — not a declared type. Sorted by identifier. This table needs no JavaScript.

Educational/research dataset. Every value traces to a published source; per-series provenance is in PROVENANCE.csv and the Methodology page. No synthetic or estimated data.