Oxford’s humanities AI project starts with the history of its training texts

Oxford’s BodleianLLM project is setting out to train a humanities language model on the university’s digitised cultural collections. Funded by the John Fell Fund for September 2026 to August 2027, it plans to release the model, training code, data documentation and tests. The project is led by Glenn Roe, with Megan Gooch and research software engineer Miguel Arana-Catania.[1]

The revealing detail is the job of preparing the collection. Oxford’s recruitment notice for the project’s postdoctoral researcher specified curating, cleaning and documenting a multilingual corpus, then continuing the pre-training and fine-tuning of an existing open-weight foundation model on university infrastructure. The proposed route therefore starts with an existing model and adapts it to scholarly material.[2]

A library collection contains decisions made long before anyone considers training AI on it. Which books survived? Which editions were collected? Which pages were photographed, transcribed or corrected? Those choices determine what a model can encounter. Recording them makes the eventual model easier to understand as a research instrument, even before its answers are assessed.

A good transcription still has a selection history

The Early English Books Online Text Creation Partnership offers a useful historical example. Its editors describe making searchable, encoded transcriptions from early printed books. Initial selection favoured named authors in a major bibliography; later efforts deliberately widened the range through subjects such as food, piracy and witchcraft, and formats including pamphlets. First editions were generally preferred. Difficult-to-read works and texts offering little verbal material were among those avoided.[3]

These choices served an expensive scholarly undertaking with finite resources. A corpus built that way can be carefully edited while representing some forms of writing much more fully than others. Accuracy within a transcription and representativeness across a historical period are different properties. Both matter when the collection becomes training material.

EEBO-TCP is context for the problem, not a confirmed BodleianLLM training set. The project page does not yet identify its final corpus, base model or model size. Its stated research tasks include manuscript transcription, recognising historical names and comparing literature across languages. Those aims describe the work ahead; the page reports no completed model performance.[1]

Making the collection usable by somebody else

Oxford has already been examining this layer of digital scholarship. A separate Bodleian–Berlin State Library collaboration, scheduled from August 2025 to April 2026, investigated how different users want to access cultural-heritage datasets and what support they need. Its questions included the treatment of sensitive or contested material, alongside the more practical matter of what formats and services people could use.[4]

The distinction between having access and being able to reuse material also appears in the Bodleian’s current text-and-data-mining guidance. It describes different arrangements for subscribed journals, institution-wide tools and resources licensed to a defined research project, with librarians helping researchers establish suitable access. A library’s holdings are not one undifferentiated pool of freely transferable text.[5]

For an open training pipeline, the documentation has to travel with the useful material. Another institution needs to know what was included, what was removed, how passages were processed and which version of a source was used. Otherwise it can copy the machinery without being able to explain the instrument it has built.

The planned release creates an opportunity to preserve scholarly choices alongside the model. A researcher studying a particular period could then ask a precise question about an answer: does the training collection contain the relevant language, edition or kind of document? That is a more useful starting point than treating a fluent response as evidence that the historical record was adequately represented.

BodleianLLM’s proposed legacy is therefore larger than a set of model weights. If the team delivers the documented training process it promises, other libraries could adapt the method to collections with different languages, periods and editorial histories. The transferable resource would include a way of showing how a model acquired its view of the past.

Report an error or suggest a correction