Rendered BookCorpus Images from 17,868 Books

Encompassing 17,868 books from the original BookCorpus, with text examples rendered as images at a resolution of 16x8464 pixels. It was created by Team-PIXEL and last updated in August 2022.

Use Cases

• Train vision-language models on rendered text images from 17,868 books to understand document layout.
• Fine-tune optical character recognition (OCR) systems on the 16x8464 pixel image format.
• Analyze text rendering consistency across different books and genres within the corpus.

Strengths

• Derived from the established BookCorpus, which contains 17,868 books.
• Provides a unique multimodal format with text rendered as fixed-resolution images (16x8464 pixels).
• Based on a corpus introduced in a seminal 2015 research paper (Zhu et al.).

Limitations

• The specific number of rendered image examples (rows) is unknown.
• Image resolution is fixed at an unconventional 16x8464 pixels, which may not be suitable for all vision tasks.
• Dataset details like license, file formats, and exact update frequency are unspecified.

Provenance

Source

Derived from the BookCorpusOpen dataset on Hugging Face, originally introduced by Zhu et al. (2015).

Collection Method

Text from the BookCorpus was rendered into images.

Freshness

Last updated on 2022-08-03.

Geography

Region tag suggests 'us', but specific geographic coverage of the source books is unknown.