Dataset

Corpus composition

The R3 synthetic corpus contains 1,350 documents in six languages: German, English, Spanish, French, Italian and Dutch.

Document families cover correspondence, minutes and decisions, contracts and procurement, personal records, finance, forms and registers, and compound documents. Jurisdictions cover Germany, Spain, EU institutions, France, Ireland, Italy, the Netherlands and the United Kingdom.

Label categories

Personal data must be redacted. Must-not-redact values represent public information that should remain visible. Policy-dependent values cover officials acting in their official capacity.

Private by design

The documents and labels stay private to protect the evaluation from memorisation and tuning to the test set. Public results contain aggregates and corpus versions; submissions receive no access to any split.