Cianmcnally/irish-census

Irish Census 1901 & 1926

Person-level records from the 1901 and 1926 censuses of Ireland, as published by the National Archives of Ireland — every individual return, in flat CSV.

Year	Rows	Size	Coverage
1901	4,434,939	4.31 GB	All of Ireland (32 counties)
1926	2,973,480	0.56 GB	Saorstát Éireann (26 counties)
Total	7,408,419	4.87 GB

The 1926 census is the first taken by the Irish Free State and was released to the public in 2026 under the 100-year rule. The 1922 Four Courts fire destroyed the 19th-century returns, which is why 1901 is among the earliest surviving full-population censuses for Ireland. Between them these two snapshots bracket the revolutionary period: the same townlands, often the same families, 25 years and one partition apart.

Usage

from datasets import load_dataset

ds_1901 = load_dataset("Cianmcnally/irish-census", "1901", split="train")
ds_1926 = load_dataset("Cianmcnally/irish-census", "1926", split="train")
both    = load_dataset("Cianmcnally/irish-census", split="train")

Columns

One row per person. year distinguishes the two censuses; a_id is the archive's record id; image_group groups the members of one household.

Shared: a_id, county, townland, ded (district electoral division), first_name, surname, relationship_to_head, updated_relationship_to_head, updated_sex, updated_marriage, updated_age, updated_religion, updated_irish_language, years_married, irish_or_english, birthplace_county, children_born_alive, children_living, image_group, year.

1901 only: house_number, education, occupation, occupation_updated, deafdumb, images.

1926 only: aform_name, geocode, institution_name, institution_type.

Columns absent for a given year are empty in that year's rows.

• updated_* columns are the archive's normalised readings of the enumerator's handwriting; the un-prefixed columns are closer to the form as written.
• a_id, updated_age, image_group and year are numeric or empty. Where the source has a non-numeric value in a numeric field ("abt 40" for an age), the cell is empty rather than coerced to a wrong number.
• images (1901) is a JSON string: the household's scan pages as [{"form","side","id","url"}], with paths relative to https://www.nationalarchives.ie. The same list repeats for every member of a household — deduplicate on image_group if you need one row per scan. It accounts for most of the 1901 file size and was kept deliberately, so the dataset links back to the original handwriting.

How it was scraped

Both censuses sit behind the National Archives' public search APIs — the same endpoints their website's search box calls:

Year	Endpoint	Record ids
1901	api-census.nationalarchives.ie/census/query	4,382,145 – 8,817,869
1926	c26-api.nationalarchives.ie/api/census/query_c26a	from 1, until the API runs dry

The scraper is a single Python script (requests + huggingface_hub, nothing else) that walks each year by record id: request 1,000 rows sorted by id with id >= N, take the highest id returned, ask for the next 1,000 from there. Keying on the id rather than page offsets makes the crawl deterministic and restartable — there is no cursor to lose.

A few design choices worth knowing about:

• No local database. Rows are buffered in memory, written out as a CSV shard of 100k–200k rows, uploaded straight to this repo, and deleted. Peak local disk during the whole scrape was one shard (~30–130 MB).
• The repo is the resume state. Shards are named by the id range they cover (data/1901/004382145-004482158.csv). On startup the script lists the repo's files, takes the highest uploaded id for that year, and continues from the next one. Kill the process, wipe the machine, run it again — it picks up where the repo says it stopped, and shard ranges stay contiguous and non-overlapping.
• Polite pacing. A random 2–4 s delay between requests, exponential backoff on errors, and a 60 s-and-growing wait on HTTP 429/503. Each year took a few hours of wall clock; the archive's API was never hammered.
• Clean shutdown. SIGINT/SIGTERM finish the in-flight batch and upload the partial shard before exiting, so no buffered rows are lost.
• Verified transfer. Field mapping, numeric coercion, CSV quoting (names like O'Brien, Jr and the comma-riddled JSON image blobs), and the resume-from-shard-names logic are covered by an offline self-test that ran before every scrape.

Both years completed with zero unrecovered errors — the only incidents were a handful of transient connection drops, which the backoff absorbed without leaving a gap.

Completeness

• 1901 record ids span 4,382,145–8,817,869; 786 ids in that range return no record from the API and are simply absent here too.
• The 1926 row count is ~1,000 above the total the archive's own search page reports; the extra rows come from the API itself.
• Transcriptions are the archive's, including their errors. Names, ages and places in historical census returns are frequently misread or mis-stated — treat any single record as evidence rather than fact.

Licence and attribution

Records are the work of the National Archives of Ireland, who make them freely available to the public. This dataset is a mirror for research convenience and claims no ownership of the underlying records. Check the archive's terms of use before redistributing or building a commercial product on it, and credit the National Archives of Ireland in anything derived from it.