Public datasets
Archival corpora that researchers have chosen to publish. Every dataset carries a licence, a plain-language note describing the models that produced its results, and a stable URL you can cite.
No datasets have been published yet. Researchers publish from their workspace, once a corpus has been processed and given a licence.
What a published dataset contains
A dataset on Archēglyph is a corpus of archival documents that has been through the full pipeline: layout assessed, text extracted by OCR or a vision-language model, split into fragments, embedded, indexed both lexically and semantically, and grouped into topical clusters. Publishing it makes four things public — the documents it contains, the clusters found in it, the versions saved from it, and a plain-language note describing exactly which models produced each of those results.
Why every result names its engine
Extracted text is not the same thing as the source. A vision-language model reading a water-damaged page produces a plausible transcription, and plausible is not accurate. So every derived block of text here carries the identifier of the model that produced it, and each section carries a note explaining how it was made. That is what makes a result arguable rather than merely presented.
Licences and citation
A dataset cannot be made public without a licence, and the licence is shown on every page and in the structured data crawlers read. Where the licence is "all rights reserved", the dataset is discoverable and describable but its contents are not yours to reproduce.
For anything you intend to cite, cite a version rather than the dataset. A version is a frozen snapshot — the text, the indices and the clusters exactly as they stood. Re-running the pipeline with a different embedding model changes the dataset and its clusters; it does not change a saved version. That is the difference between a reference that still resolves in two years and one that quietly means something else.