The War Atlas open dataset: 3,010 battles across 493 wars

Structured facts behind The War Atlas: battles with dates, coordinates and their precision, force-strength and casualty estimates, commanders, factions, and 19,020 bibliographic source records. Plain CSV plus one nested JSON file, no API key, CC BY 4.0.

This card says what is in the files and also what is missing. Every number below was computed from the files by scripts/check_dataset.py, and CI re-runs that check on every push.

• 3,010 battles in 325 wars (493 wars are listed; 168 have no battles yet), dated 1763 to 2026.
• 2,477 battles with coordinates (82.3%), each tagged exact or place-level, with where the point came from.
• 6,618 force and casualty estimates (3,905 + 2,713), stored as low and high.
• 12,911 battle-to-commander links and 9,886 battle-to-source links in battles.json.

The one idea worth stealing, even if you never use this data: publish the coverage next to the count. "3,010 battles with source-cited casualty ranges" sounds complete; "1,923 battles have a casualty estimate, and 45 of those rows name their own source" is what an analyst needs before trusting a chart. A dataset card that states its gaps is the one people can build on.

---

60 seconds to a first query

No clone needed. Python with pandas reads straight from GitHub (examples/quickstart.py):

pip install pandas
curl -sO https://raw.githubusercontent.com/Purcell-Analytics/war_atlas_data/main/examples/quickstart.py
python quickstart.py

import os

import pandas as pd

# Reads straight from GitHub. Set WAR_ATLAS_BASE to a local folder (ending in /) to use a clone.
BASE = os.environ.get("WAR_ATLAS_BASE", "https://raw.githubusercontent.com/Purcell-Analytics/war_atlas_data/main/")


def load(name):
    # keep_default_na=False: only a truly empty cell is missing, so a side named "NA" stays a string.
    return pd.read_csv(BASE + name, keep_default_na=False, na_values=[""])


battles, forces, wars = load("battles.csv"), load("battle_forces.csv"), load("wars.csv")

Real output (captured in docs/assets/src/captures/quickstart_py.json):

3,010 battles, 493 wars listed, 325 wars with at least one battle
2,477 battles have coordinates: {'exact': 2052, 'place': 425}
                    title  year           side  strength_low  strength_high
         Battles of Rzhev  1942   Soviet Union       3680300        3680300
            Great Retreat  1915 Russian Empire       2975695        2975695
    Battle of the Dnieper  1943   Soviet Union       2633000        2633000
Battle of the Oder–Neisse  1945   Soviet Union       2500000        2500000
            Great Retreat  1915  German Empire       2411353        2411353

JavaScript (Node 18+, no dependencies) loads the nested battles.json and turns it into GeoJSON (examples/quickstart.mjs):

curl -sO https://raw.githubusercontent.com/Purcell-Analytics/war_atlas_data/main/examples/quickstart.mjs
node quickstart.mjs

3010 battles, 2477 as GeoJSON points
Battle of Ayacucho 1824 1824-12-09
  force  Patriots: 5780-8500
  force  Royalists: 6906-9310
  losses Royalists: 2500-2500
  4 commanders, 1 source references

Both examples run in CI against the checked-out files on every push.

Files

File	Rows	One row per
battles.csv	3,010	battle
battle_forces.csv	3,905	force-strength estimate for one side of one battle
battle_casualties.csv	2,713	casualty estimate for one side of one battle
wars.csv	493	war
commanders.csv	8,012	commander (mostly people; some rows are units, e.g. 15th_army_group)
factions.csv	425	faction
sources.csv	19,020	bibliographic source record
battles.json	3,010	battle, with nested forces[], casualties[], commanders[], source_slugs[]
manifest.json		snapshot version (2026-07-03), row counts, license, citation
dataset-metadata.json		Kaggle dataset metadata

Schema

Every table is keyed by a slug. "Filled" is the share of rows where the field is non-empty. Types are as they appear in the CSV; an empty cell means unknown.

battles.csv

Field	Type	Filled	Notes
slug	string, unique	100%	primary key, e.g. battle_of_ayacucho_b
title	string	100%	English title; not unique (see limitations)
war_slug	string, FK wars.slug	100%
campaign_slug	string	3.2%
year	integer	100%
date_iso	YYYY-MM-DD	98.5%	year always equals year; 261 are YYYY-01-01 (see limitations)
date_display	string	100%	human-readable date or range
location	string	1.5%
lat, lng	decimal degrees	82.3%	both or neither; all in range; none at (0, 0)
coord_precision	exact | place	82.3%	exact: 2,052, place: 425
coord_source	battle_p625 | located_in	82.3%	Wikidata coordinate (P625) of the battle, or of the place it is located in
result	string	1.5%
intensity	integer 0 to 100	100%	site-internal score; 1,011 battles are 0
confidence_level	low | medium | high	1.5%
is_indexable	True | False	100%	site page indexing flag; 61 are False
wikidata_qid	Q + digits, unique	100%
wikipedia_url	URL	100%	built from the title (see limitations)
url	URL	100%	page on thewaratlas.co (redirects to a hyphenated path)

battle_forces.csv and battle_casualties.csv (same shape)

Field	Type	Filled	Notes
battle_slug	string, FK battles.slug	100%	one row per (battle, side), checked unique
side	string	100% (1 empty)	free text, mixed languages (see limitations)
strength_low, strength_high / casualties_low, casualties_high	integer	100%	0 <= low <= high holds on every row
confidence	low | medium | high	100%	every row in this snapshot is medium
source_slug	string, FK sources.slug	1.6% / 1.7%	64 force rows and 45 casualty rows

wars.csv

Field	Type	Filled	Notes
slug	string, unique	100%
title	string	100%
era_slug	string	100%	8 eras, e.g. cold_war (99 wars), founding (85)
period	string	100%	display text, e.g. 1808-1833
start_year, end_year	integer	100%	start_year <= end_year on every row
status	scaffolded | modeled	100%	461 scaffolded, 32 modeled
wikidata_qid	Q + digits	100%
url	URL	100%

commanders.csv

Field	Type	Filled	Notes
slug	string, unique	100%	7,646 of 8,012 are linked from a battle
name	string	100%
primary_faction_slug	string, FK factions.slug	15.7%
birth_year, death_year	integer	38.3% / 36.5%
wikidata_qid	Q + digits	39.7%
url	URL	100%

factions.csv: slug (unique, 100%), name (100%), wikidata_qid (39.8%), url (100%).

sources.csv: slug (unique, 100%), source_type (100%: 14,867 monograph, 4,142 academic_journal, 11 others), title (100%), url (58.5%). 10,903 slugs are ISBN-keyed (e.g. isbn-0333418379); the rest are slugified titles.

battles.json: an array of the same 3,010 battles with every battles.csv field (numbers typed, lat/lng null when unknown) plus forces[], casualties[], commanders[] (slug, name, side, rank) and source_slugs[]. The checker confirms it agrees with the CSVs field for field.

Sample rows

One battle traced through every file (from captures/sample_rows.json):

battles.csv            battle_of_ayacucho_b | Battle of Ayacucho | war spanish_american_wars_of_independence
                       1824 | 1824-12-09 | -13.0425, -74.131667 | exact | battle_p625 | Q934157
battle_forces.csv      Patriots  5780-8500  medium | Royalists  6906-9310  medium
battle_casualties.csv  Royalists 2500-2500  medium
wars.csv               spanish_american_wars_of_independence | founding | 1808-1833 | scaffolded | Q1123201
commanders.csv         antonio_de_sucre | Antonio de Sucre | 1795-1830 | Q189779
factions.csv           peru_bolivian_confederation | Peru–Bolivian Confederation
sources.csv            quinto-centenario-1985 | academic_journal | Quinto Centenario | dialnet.unirioja.es

Coverage

	Count	Share
Battles with coordinates	2,477 of 3,010	82.3%
Battles with at least one commander	2,562 of 3,010	85.1%
Battles with a force estimate	2,229 of 3,010	74.1%
Battles with a casualty estimate	1,923 of 3,010	63.9%
Battles with either estimate	2,475 of 3,010	82.2%
Battles with at least one source reference	1,379 of 3,010	45.8%
Force rows where low < high (a real range)	697 of 3,905	17.8%
Casualty rows where low < high	601 of 2,713	22.2%
Force rows naming their own source	64 of 3,905	1.6%
Casualty rows naming their own source	45 of 2,713	1.7%
Wars with at least one battle	325 of 493	65.9%

Regenerate all of it with python3 scripts/check_dataset.py --stats.

Provenance and sourcing method

• Entities and coordinates. Battle, war, commander and faction records are normalized largely from Wikidata (CC0) and cross-linked by slug. Every battle and war carries its Wikidata QID. coord_source says whether the point is the battle's own Wikidata coordinate (battle_p625, 2,052 battles) or the coordinate of the place it is located in (located_in, 425).
• Force and casualty figures. Harvested from Wikipedia infoboxes and kept as low/high integers per side. Battles with no figure have no row; nothing is imputed.
• Sources. sources.csv holds bibliographic records (mostly ISBN-keyed monographs and journal articles). They are linked to battles at the battle level through battles.json source_slugs[]. Only 109 estimate rows link a source directly, and those point to a handful of US archive collections (Library of Congress, NARA, NPS, ABMC, VA).
• No prose. Harvested Wikipedia text (CC BY-SA) is excluded on purpose, so the compilation can be licensed CC BY 4.0. There are no description columns.
• Snapshot. This repo mirrors the files published at thewaratlas.co/data as of version 2026-07-03 (see manifest.json). The site regenerates its copy as the catalogue grows, so it may be newer.

Known limitations

Read these before drawing conclusions. Counts are from --stats on this snapshot.

1. Most estimates are not row-level cited. 64 of 3,905 force rows and 45 of 2,713 casualty rows have a source_slug. Battle-level references exist for 1,379 battles, but they are not tied to a specific number.
2. confidence carries no signal in the estimate files. Every force and casualty row is medium.
3. Most "ranges" are single values. 3,208 of 3,905 force rows and 2,112 of 2,713 casualty rows have low == high.
4. wikipedia_url is built from the title, not taken from Wikidata sitelinks (2,969 of 3,010 equal https://en.wikipedia.org/wiki/ + title). Same-named battles share a URL: 83 URLs are shared by 184 battles (e.g. the Battle of Bautzen of 1813 and of 1945), so at least one in each group points to the wrong or a disambiguation page. Use wikidata_qid to join to Wikipedia.
5. Year-only dates look like January 1. 261 battles have date_iso ending in -01-01, and date_display repeats it (e.g. the 1842 retreat from Kabul shows "January 1, 1842"). Treat -01-01 as year precision unless you have checked it.
6. War assignment is not always within the war's years. 320 battles have a year outside their war's start_year..end_year. Some are framing (WWII is stored as 1941–1945 U.S. combat role), some look misassigned (a 1952 clash filed under the 1882 Anglo-Egyptian War).
7. side is free text in several languages ("Royaume-Uni", "Югославия", "United Kingdom") and is not linked to factions.csv. Normalize before grouping by side.
8. Not every row is a fought battle, and duplicates are possible. For example seven_days_to_the_river_rhine_b is a 1979 Warsaw Pact exercise plan and carries a 2,000,000 casualty figure. 5 title-and-year pairs appear twice with different QIDs (e.g. Battle of Arica, 1880), and 3 Wikidata QIDs are shared by two factions (e.g. east_india_company and british_east_india_company). The checker does not adjudicate history.
9. Sparse curated fields. result, location and confidence_level exist for 44 battles; campaign_slug for 96.
10. Coverage is uneven. Located battles cluster in Europe and North America (see the map in the hero image), and the war eras follow US-history periods (founding, continental_expansion, civil_war).

Integrity checks

scripts/check_dataset.py (standard library only) runs 79 named checks and exits 1 on any failure:

Check family	What fails it
rows.match_manifest, rows.floor	a file whose row count disagrees with manifest.json, or drops below this snapshot's count
schema.header, required	a renamed or missing column; an empty required field
types.*	a non-integer year, intensity outside 0..100, a malformed QID or date, a date whose year disagrees with year
coords.*	lat without lng, out-of-range coordinates, (0, 0), precision without coordinates
unique.*	duplicate slugs, duplicate battle QIDs, two estimates for the same (battle, side)
fk.*	a battle pointing at a missing war, an estimate at a missing battle or source, a commander at a missing faction, a JSON link to a missing commander or source
range.*	low > high, negative values, an unknown confidence value
json.*	battles.json disagreeing with any CSV field, force row or casualty row

python3 scripts/check_dataset.py           # 79/79 checks passed
python3 -m unittest discover -s tests -v   # 23 tests: each check proven able to fail

CI runs both, fails if fewer than 23 tests ran, then runs the two loading examples.

Scope: what this dataset is not

• Not a casualty authority. The figures are infobox estimates, mostly uncited at row level. Do not publish a total from them without going back to sources.
• Not complete. It covers the battles The War Atlas has catalogued, weighted toward the eras the site started with. Absence of a battle means nothing.
• Not reconciled across languages. Side names are kept as harvested.
• Not live. A versioned snapshot; the site copy may be newer.

The patterns

Pattern	The failure it prevents
Coverage published next to every count	"3,010 battles with casualty ranges" read as 3,010 battles with casualty figures
low/high instead of one number	a contested figure collapsed into false precision
coord_precision + coord_source on every point	a city centroid plotted as if it were the battlefield
No row for an unknown figure	an imputed or zero value summed into a total
Row-count floors in CI	a truncated export that still parses and quietly ships
JSON checked against CSV	two formats of the same data drifting apart

FAQ

Can I use this commercially? Yes, under CC BY 4.0, with attribution (below).

Why do some battles have no coordinates? 533 battles have no usable Wikidata coordinate for the battle or its location. They are left empty rather than guessed.

Why do numbers on the site differ from this repo? The site regenerates its export as the catalogue grows; this repo is the 2026-07-03 snapshot.

How do I join to Wikipedia reliably? Use wikidata_qid, not wikipedia_url (limitation 4).

Citation

The War Atlas (2026). The War Atlas Open Dataset (version 2026-07-03) [Data set].
https://thewaratlas.co/data

@misc{thewaratlas_2026,
  title        = {The War Atlas Open Dataset},
  author       = {{The War Atlas}},
  year         = {2026},
  version      = {2026-07-03},
  howpublished = {\url{https://thewaratlas.co/data}},
  note         = {CC BY 4.0. Mirror: https://github.com/Purcell-Analytics/war_atlas_data}
}

License

CC BY 4.0 (full text in LICENSE). You may share and adapt the data for any purpose, including commercially, if you give attribution, link to the license, and indicate changes. Suggested credit:

Data: The War Atlas, thewaratlas.co, CC BY 4.0

Underlying facts are largely derived from Wikidata (CC0); the compilation, selection and curation are licensed CC BY 4.0. The scripts and examples in this repo are covered by the same license.