The xd Crossword Corpus

AI and the corpus

Draft, November 2026, for a corpus begun in 2016 (see HISTORY.md). Marks as in GOVERNANCE.md.

What the corpus is

The corpus holds puzzles made by people, by named constructors where the name is known, as the publisher meant them: obvious misprints corrected and recorded, nothing deliberate changed. It does not collect puzzles generated by a program or an AI with no one constructing or editing them (see COLLECTION.md, “Made by a person”). It does not generate puzzles. [established]

How we use AI ourselves

OCR models read page scans of old puzzles. Language-model tools help us maintain the corpus: converting and checking files, reviewing changes, and drafting documents, including this one. They do not compose puzzles or clues; where a puzzle has no key, the answers may be solved from the clues, and the source is recorded in receipts. Nothing they produce enters the corpus without a person reviewing the change before it is merged. [practice]

Using the data to train models

The tiers are those of DATASETS.md.

• Public-domain puzzles. Anyone may use them for anything, including training models. We cannot restrict them and do not pretend to.
• Public derived data, such as the clue and answer file. Our terms ask that it be used for research, analysis, tools and solving, not to build systems that generate crosswords or clues for publication, and that the corpus be cited. We attach these terms to the data as signals: fields on the Hugging Face dataset card, and machine-readable training opt-outs where the venue supports them. They are terms we ask people to honor, not locks. [proposed]
• Gated samples and the full corpus. The terms in ACCESS.md apply, which forbid using them to deliver puzzles to solvers. We add the same term: not to build systems that generate crosswords or clues for publication. [proposed]

The clue file has been public for years, and has already been used to train crossword-solving models.

Constructors and publishers

Removal on request, as described in CONTRIBUTING.md, applies here too. A constructor who does not want their clues in the public derived data can ask to be left out of future snapshots, and we will do so. [proposed]

Review

This statement is reviewed once a year with the other policies. It is a first version. We want to hear what constructors, editors, solvers and researchers think of it: open an issue in this repository. Changes are logged in DECISIONS.md.

This site is open source. Improve this page.