SchoolData/us-k12-schools

US K–12 Schools — Open Dataset for the AI Era

Encyclopedic text for every one of the 122,476 K–12 schools in the United States, ready to use as pretraining text.

Information about American schools is fragmented and unevenly represented in the public sources that language models learn from. Only about 12.4% of schools have a Wikipedia article, and most of those are high schools; for most others, the main public record is a federal directory listing. This dataset gives every school one consistent, sourced text record: the school's Wikipedia article where a verified one exists (12.4% of rows, once unreliable matches are excluded), and otherwise a paragraph rendered deterministically from public federal data. Every row says where its text came from.

from datasets import load_dataset
ds = load_dataset("SchoolData/us-k12-schools", "articles", split="train")

Dataset Summary

Rows (one per school)	122,476
from_wikipedia = 1 (Wikipedia text)	15,163 (12.4%)
from_wikipedia = 0 (generated from NCES)	107,313 (87.6%)
Public / private	100,237 / 22,239
States and territories	56
Total text	~98.4M characters

A single config, articles. One row per school, keyed by NCES school ID. Every row has a text field; from_wikipedia tells you where that text came from.

Data Fields

Field	Type	Description
nces_id	string	NCES school ID — the primary key, unique across the dataset
name	string	School name as recorded by NCES
state	string	Two-letter state or territory code (where the school is located)
sector	string	public (CCD) or private (PSS)
text	string	The school's encyclopedic text
from_wikipedia	int8	1 = text from Wikipedia, 0 = generated from NCES
wikipedia_title	string	Source article title (from_wikipedia = 1 only)
wikidata_qid	string	Wikidata item for the school (from_wikipedia = 1 only)
wikipedia_revid	string	Exact source revision id (from_wikipedia = 1 only)
source	string	Article URL, or NCES CCD 2024-25 / NCES PSS 2023-24
license	string	CC-BY-SA-4.0 or CC0-1.0 — see Licensing

Licensing

The dataset as a whole is distributed under CC BY-SA 4.0.

It mixes two license regimes, and the per-row license field records which applies:

Rows	Origin	Row license
from_wikipedia = 1	English Wikipedia	CC-BY-SA-4.0
from_wikipedia = 0	NCES CCD / PSS — US federal works, public domain	CC0-1.0

Because Wikipedia text is included, ShareAlike propagates to the combined distribution, so the whole dataset ships as CC BY-SA 4.0. If you need a permissively licensed subset, filter it out:

permissive = ds.filter(lambda r: r["from_wikipedia"] == 0)   # 107,313 rows, CC0-1.0

Wikipedia Attribution

Rows with from_wikipedia = 1 contain text derived from English Wikipedia, licensed CC BY-SA 4.0. Each such row links to its source article via wikipedia_title / wikidata_qid / wikipedia_revid; authorship is the article's revision history. Text was extracted and cleaned — it is the article's prose, not its verbatim wikitext. This dataset is redistributed under CC BY-SA 4.0.

Concretely, the changes made to Wikipedia content are: the full article body was taken as plain text via the MediaWiki extracts API (infoboxes, tables and references are not part of the extract); the References, External links, See also, Notes, Further reading and similar trailing sections were dropped; section headings were removed and whitespace collapsed to single spaces; and surviving markup — unexpanded {{templates}}, raw [[wikilinks]], and inline [12] citation markers — was stripped. Wikipedia content was retrieved on 2026-10-11; wikipedia_revid pins the exact revision each row was taken from.

Non-Wikipedia rows

The 107,313 rows with from_wikipedia = 0 are rendered from NCES public-domain data by a deterministic template — no language model, no randomness, no inference. Every clause is optional and appears only when the underlying field exists, so a missing value produces a shorter sentence rather than a guess. Nothing in these rows is invented.

Limitations & Biases

Wikipedia coverage is unevenly distributed, and that is the problem this project exists to document. Among states with at least 100 schools, coverage ranges from 0.6% (Puerto Rico), 5.3% (South Dakota) and 5.7% (Montana) up to 20.3% (Oregon) and 22.6% (Arkansas). Coverage also differs sharply by level: most articles are about high schools. How coverage varies with locale and school poverty has not been measured yet, so this card makes no claim about it. Training on from_wikipedia = 1 alone would inherit whatever skew exists.

Generated rows are thin by construction. They contain only what NCES records — location, grade range, district, enrollment, staffing, and for private schools religious affiliation, demographics and facilities. No history, no notable alumni, no local context. They establish that a school exists and state verifiable facts about it; they are not a substitute for a written article.

Wikipedia matching is imperfect. Articles are linked to NCES schools by Wikidata NCES ID (11,211 matches, plus 69 through re-issued ids), exact name + state (2,605), fuzzy name similarity within a state (699), or the nearest NCES school within 1 km sharing a name word (580). Only articles verified as schools are matched: people, events, districts, networks, colleges, closed schools and titles that merely redirect to a town page are excluded before matching. Every name-based match must be consistent with the title's city and with the school's NCES grade span — a high-school article is never attached to a K–5 school. Articles whose first sentence says the school was something (a closed school) never supply text to a current school (0 such matches fell back to generated text, as did 1 matches with a missing or very short article). Residual risk remains where an article states no level or place; treat fuzzy- and geo-tier rows as the least reliable part of the dataset.

A few articles describe an institution larger than the school. Some Catholic parish schools link to an article covering the parish church and its school together, so the text includes material about the church. This affects a small number of rows.

The data is a snapshot. Public-school figures are from the 2024-25 CCD, private schools from the 2023-24 PSS; Wikipedia text was retrieved 2026-10-11. Schools open, close, and merge; enrollment moves.

Text length varies by source. Median 1,980 characters for Wikipedia rows (full articles), 490 for generated rows.

Provenance & Fiscal Host

Source	Contributes	Original license
NCES Common Core of Data (CCD) 2024-25 + EDGE geocodes	Public school records	US federal work, public domain
NCES Private School Survey (PSS) 2023-24	Private school records	US federal work, public domain
Wikidata	NCES ID linkage, QIDs	CC0
English Wikipedia	Article text for 15,163 schools	CC BY-SA 4.0

This is a non-profit, community-verified civic data project, fiscally sponsored by Hack Club Bank, a 501(c)(3) organization.

Reproducing this release

Every stage of this dataset — the NCES build, the Wikipedia crawl and matcher, and the release build — is open source at github.com/bxshan/SchoolData. Given the same inputs the build is deterministic (no model, no randomness); build_manifest.json in the release records the source commit and the sha256 of every input.

git clone https://github.com/bxshan/SchoolData && cd SchoolData/data
git lfs pull                                   # data CSVs are stored in Git LFS

# NCES masters (public from bulk files, private from the PSS search export)
python nces_crawl/build_from_bulk.py --download
python nces_crawl/download_schools.py --type private
python nces_crawl/combine_all_schools.py

# Wikipedia: crawl -> enrich -> match -> article text
python wiki_crawl/crawl_k12_schools.py --include-defunct
python wiki_crawl/enrich_schools.py
python data_publish/match_wiki_nces.py
python wiki_crawl/fetch_article_text.py --matches data_publish/output/wiki_nces_matches.csv

# Release
python data_publish/prep_data_publish.py       # renders NCES text; -> output/dist/
python data_publish/validate_publish.py --expect-rows 120000

prep_data_publish.py --state CA builds a single-state slice; --skip-match builds the NCES-only subset with no Wikipedia join. See CHANGELOG.md for the inputs behind each version.

Citation

@misc{us_k12_schools,
  title  = {US K--12 Schools: An Open Dataset for the AI Era},
  author = {Shan, Boxuan and {The School Data Project}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/SchoolData/us-k12-schools},
  note   = {CC BY-SA 4.0. Contains text from English Wikipedia.}
}