British Library Book Images

The corpus is a public-domain, large-scale collection of book-derived images that is simultaneously a convenient multimodal resource and a highly non-representative record of what a major library digitised in the early 2010s. The key practical insight is: it works well as a free, ready-made image–text collection (including precomputed embeddings), but models trained on the whole set will overwhelmingly reflect late‑Victorian illustration styles and the selection biases of the digitisation programme.

What Sets It Apart

• Size and provenance: over one million cropped images from ~65,000 volumes digitised by a national library programme; images originate from a 2014 British Library release and are redistributed here in parquet form. This makes large-scale image+OCR experiments possible without copyright friction, since the images are public-domain.
• Algorithmic image-type split: images are provided in four configs (embellishments, plates, medium, covers). the labels are the result of heuristic cropping, not curator-reviewed art-historical categories — expect fuzziness at class boundaries and treat these labels as size/position heuristics.
• Ready-made embeddings: a SigLIP2 embedding split covers every image, enabling text-to-image nearest-neighbour search and scalable retrieval experiments without recomputing encodings.
• Joinable to OCR: filenames embed the British Library system number so most images can be linked at book level to the OCR text corpus (biglam/blbooks-parquet), enabling image–text alignment and captioning tasks.

Who It's For and Trade-offs

Great fit if you need a large, public-domain multimodal corpus for retrieval, zero-shot image-to-text experiments, image-caption fine-tuning, or digital-humanities analysis of printed illustration trends. The dataset is especially useful when you want quick access to embeddings or to reproduce historical-illustration–heavy model behaviour.

Look elsewhere if you require a demographically or temporally representative sample of printed imagery, curated art-historical labels, or guaranteed page-level alignment between image crops and OCR text. Be aware of these trade-offs: the collection is dominated by 19th-century illustrations (the 1890s alone are ~33% of images), contains colonial-era and period-specific depictions that have not been content-moderated, and includes some catalogue/date errors. Also note the mirror status: this Hugging Face packaging is a static repackaging of the original deposit and receives limited maintenance.

Practical notes

• Data shape: images at source JPEG resolution; metadata fields include date (string, some "Unknown" and a few erroneous 20th-century dates), fname (filename, leading digits = system number), and image_type.
• Scale and access: the full dataset is large (hundreds of GB); streaming is recommended for exploratory work. Parquet splits map to the four image-type configs; embeddings are provided in a separate embeddings config and align row-for-row with image rows.
• Ethics and curation: images were algorithmically harvested and not reviewed for offensive content; colonial-era imagery and period captions are present. Filter and audit for downstream uses that require content safety or representativeness.