Cultural heritage institutions hold vast collections, and the effort to digitally catalogue them for research and public access has been enormous. Yet the backlog continues to grow. Many existing records were created before current standards and technologies matured, resulting in inconsistent, incomplete, or error-prone metadata. Language models offer a route to addressing this at scale but the typical paradigm of large, compute-heavy models is ill-suited for a sector with limited hardware budgets and strict obligations around copyright and data privacy. Institutions often cannot share collection data with third-party services and rarely have the infrastructure to even run models that are referred to as “small” in language modelling research. The result is a transformative technology, placed tantalisingly out of reach of a sector that stands to gain enormously from its application.
This paper investigates how language models can be practically deployed within Galleries, Libraries Archives and Museums (GLAM) digitisation and data enrichment workflows, in partnership with the National Gallery. Rather than pursuing scale, it promotes the fine-tuning of smaller models for specific tasks on institutional data. The work identifies several pivotal tasks involved in producing quality data for heritage research and public dissemination where this proves especially valuable: correcting errors introduced during digitisation, standardising inconsistent metadata, extracting structured information from unstructured text and disambiguating references to people, places, and concepts within collections records.
To keep deployment feasible under real institutional constraints, the research makes use of contemporary efficiency techniques such as quantisation and Low-Rank Adaptation (LoRA). By sharing a single base model and attaching lightweight, task-specific adapters, multiple specialised models can be served without substantially increasing memory requirements.
Fine-tuning demands training data that many institutional tasks lack. In such cases, synthetic data generation becomes a practical necessity. The research examines strategies for producing synthetic data of a higher quality, that are more challenging and more representative of the characteristics present in heritage collections. In particular, it investigates the counterintuitive notion of introducing noise into evaluation datasets to better emulate the reality of typical source material.
A further contribution addresses what is arguably the most significant barrier to responsible adoption of language models in the sector: the absence of domain-specific evaluation benchmarks. Without rigorous ways to measure whether a model actually performs well on heritage tasks, institutions have little basis for informed decision-making. This research constructs several benchmark datasets by leveraging the structure of existing collections metadata to derive labels, testing domain knowledge and performance across common sectoral tasks while leveraging the expense, effort and expert knowledge already invested into institutional collections.
As a primary case study, the research draws on real user search queries from the National Gallery's discovery endpoint to evaluate embedding models, providing a grounded measure of retrieval quality. This embedding evaluation is presented through an interactive dashboard, offering a more visual and accessible evaluation methodology.
Taken together, the result is a pragmatic approach to integrate language models into heritage workflows: efficient models that run within institutional constraints, methods for faithfully generating the training data those models need, and evaluation tools to assess whether they are actually working. The aim throughout is to move language models from a promising but inaccessible technology toward something heritage institutions can meaningfully adopt and critically assess.