The xd Crossword Corpus

Datasets

Draft, November 2026. Marked [practice] where it describes what we already publish, [proposed] where it is new.

Purpose

Most puzzles in the corpus are under copyright, and we do not publish them (see ACCESS.md). But much work with crosswords does not need whole puzzles: it needs the words, the clues, who edited what and when. The public data is for anyone who wants to study crosswords or build something with crossword data: researchers, and just as much hobbyist programmers, since the corpus and most of the crossword world are hobby projects. We publish it so that this can happen without anyone having to ask, and without redistributing anyone’s puzzles.

The principle behind every public dataset: nothing we publish lets a copyrighted puzzle be rebuilt.

Three tiers

(a) Public full content

Complete puzzles that are out of copyright:

• Before 1931: public domain in the United States outright. The corpus holds about 700: Judge magazine 1924 to 1930 (digitized by the Every Puzzle Project), the 1924 Simon & Schuster book, and a few single puzzles. [proposed]
• Judge, 1931 to 1939: about 300 puzzles. US copyright in works published 1931 to 1963 lasted only if renewed in the 28th year, and a check of the renewal records (assessments/copyright-renewal-2026.md) found no renewal of any Judge issue. [proposed]
• New York Times, 1942 to 1963: not public domain. A check of the renewal records (assessments/copyright-renewal-2026.md) found that the Times renewed the copyright in every issue of those years, so these 5,267 puzzles stay under copyright until 2037 to 2058. They belong to tier (c), the gated tier, not here. The puzzles-pre1965 release of century-arcade/xd published them on the belief that the Times had not renewed; the maintainers are deciding what to do with that release. [proposed]

Anyone can download them, from GitHub releases and Hugging Face.

(b) Public derived data

• Clues and answers (clues.tsv in xd-clues.zip): one row per clue, with publication, year, answer and clue. [practice]
• Puzzle metadata (xd-metadata.zip): puzzles.tsv (identifier, date, size, title, constructor, editor, copyright line), stats.tsv (puzzle counts by publication, year and weekday) and similar.tsv (for each puzzle, how much of its grid and clues match earlier puzzles). [practice]
• Answer lexicon by year: each answer with its number of uses per publication-year. [proposed]
• Grid statistics: size, block count, word count and word lengths, without the grid itself. [proposed]

Anyone can download them, today from xd.saul.pw/data, and from Hugging Face and Kaggle once they are set up.

(c) Gated

Samples of full modern puzzles (see “Samples” in ACCESS.md).

• A random sample of 1,000 puzzles, drawn across the whole corpus, with full .xd content. It is fixed and versioned: the same puzzles every time, so results can be compared. A researcher building a crossword solver already has it. [practice]
• Variations can follow on request: a sample stratified by decade or by publication, a date-split benchmark (older puzzles for training, recent ones for testing), and larger samples. [proposed]

These are published on Hugging Face as gated datasets: each person asks, a maintainer approves them by name, and the ACCESS.md terms apply, including the rule against delivering puzzles to solvers. This is the same process as a request by email, with a form, and hobbyists apply the same way as researchers. [proposed]

The full corpus is not a dataset. It stays available only by direct arrangement under ACCESS.md.

Rules for derived data

Tier (b) follows these rules, [proposed] unless marked.

• No grids, no clue numbers. A clue row does not say where its answer sits in a grid, or which clue it was. [practice]
• Year, not date. Clue rows carry the year of publication and nothing finer. [practice]
• Rows are reordered within each publication-year. The order of a puzzle’s clues does not survive; today the rows of each publication-year are sorted by answer and clue. [practice]
• Contest puzzles with redacted grids are left out. Their clues are not published, so a contest is not spoiled before its answers are. [practice]
• A minimum cell size. A publication-year with fewer than 5 puzzles is too small to hide which clue came from which puzzle. Its rows are merged into a coarser cell (the publication’s decade) or dropped.
• A coverage table. Coverage is very uneven, so every dataset card carries a table of puzzle counts by publication and decade.

puzzles.tsv is per puzzle and carries full dates, but no clues, and no answers beyond the first across and down, used as a grid hash. [practice]

Where and how

• Hugging Face, under one organization (name to be decided; we propose xdcorpus), holds tiers (a), (b) and the gated tier (c). [proposed]
• Kaggle, equally, holds tier (b): the same snapshots with the same card text, published at the same time as on Hugging Face. [proposed]
• Snapshots, not streams. Datasets are versioned snapshots, published quarterly, not rebuilt nightly, so that a result can name the version it used. [proposed]
• Zenodo holds one record for the corpus as a whole, with a DOI for citing it, pointing at the corpus page. [proposed]
• GitHub releases keep the public-domain puzzle sets, as they do now. [practice]

Every dataset card states the terms in one paragraph, links to this document, to ACCESS.md, and to AI.md, the corpus’s statement on AI, which sets the terms on using the data to train models; says how to cite the corpus (see README.md); and carries the coverage table. [proposed]

Publishers and constructors

Publishers and constructors can ask for their puzzles to be removed from tier (a) or tier (c), as described in CONTRIBUTING.md. The puzzles are removed from the next snapshot, and gated users are asked to delete their copies.

A tier (b) row cannot be traced to one puzzle, but a publication can ask to be left out of future snapshots, and we will do so. [proposed]

Review

This policy is reviewed once a year with the other policies and the preservation assessment. Changes are logged in DECISIONS.md.