aimgo/Acervus

Pending public release - check back on February 14, 2026.

Acervus is an extensively curated and cleaned Latin language dataset spanning all orthographies and periods.

It contains nearly 24 billion high quality Latin tokens from a wide array of sources, predominantly free of OCR artifacts.

This includes nearly 1 million novel samples recovered from poorly digitized manuscripts.

As of January 2026, it is likely the largest publicly available corpus of clean Latin text to date.

Samples are marked according to provenance and inferred quality. Organic samples were sourced primarily manually from a large number of established public Latin repositories. Corrected samples are derived from heavily damaged excerpts found within LatinPD and refined with the aid of Emendator, an OCR correction model trained expressly for Latin. Samples obtained from CommonCrawl, CommonCorpus, LatinPD, FineWeb2, and FinePDF are similarly marked.

@misc{mccarthy2025Acervus,
  author       = {McCarthy, A. M.},
  title        = {{Acervus}: A Clean Latin Text Dataset},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/aimgo/Acervus}},
  note         = {Dataset}
}

This dataset is still a work in progress. There remain many uncorrected manuscripts, but OCR error correction at scale requires potentially thousands of hours of inference. Acervus was produced without any formal institutional backing under concomitant resource restraints.