Dataset Card for Team-PIXEL/rendered-bookcorpus

Dataset Summary

This dataset is a version of the BookCorpus available at https://huggingface.co/datasets/bookcorpusopen with examples rendered as images with resolution 16x8464 pixels.

The original BookCorpus was introduced by Zhu et al. (2015) in Aligning Books and Movies: Towards Story-Like Visual Explanations by Watching Movies and Reading Books and contains 17868 books of various genres. The rendered BookCorpus was used to train the PIXEL model introduced in the paper Language Modelling with Pixels by Phillip Rust, Jonas F. Lotz, Emanuele Bugliarello, Elizabeth Salesky, Miryam de Lhoneux, and Desmond Elliott.

The BookCorpusOpen dataset was rendered book-by-book into 5.4M examples containing approximately 1.1B words in total. The dataset is stored as a collection of 162 parquet files. It was rendered using the script openly available at https://github.com/xplip/pixel/blob/main/scripts/data/prerendering/prerender_bookcorpus.py. The text renderer uses a PyGame backend and a collection of merged Google Noto Sans fonts. The PyGame backend does not support complex text layouts (e.g. ligatures and right-to-left scripts) or emoji, so occurrences of such text in the BookCorpus have not been rendered accurately. Each example consists of a "pixel_values" field which stores a 16x8464 (height, width) grayscale image containing the rendered text, and an integer value "num_patches" which stores how many image patches (when splitting the image into 529 non-overlapping patches of resolution 16x16 pixels) in the associated images contain actual text, i.e. are neither blank (fully white) nor are the fully black end-of-sequence patch.

The rendered BookCorpus can be loaded via the datasets library as follows:

from datasets import load_dataset

# Download the full dataset to disk
load_dataset("Team-PIXEL/rendered-bookcorpus", split="train")

# Stream the dataset directly from the hub
load_dataset("Team-PIXEL/rendered-bookcorpus", split="train", streaming=True)

Dataset Structure

Data Instances

• Size of downloaded dataset files: 63.58 GB
• Size of the generated dataset: 63.59 GB
• Total amount of disk used: 127.17 GB

An example of 'train' looks as follows.

{
    "pixel_values": <PIL.PngImagePlugin.PngImageFile image mode=L size=8464x16
    "num_patches": "498"
}

Data Fields

The data fields are the same among all splits.

• pixel_values: an Image feature.
• num_patches: a Value(dtype="int64") feature.

Data Splits

train
5400000

Dataset Creation

Curation Rationale

More Information Needed

Source Data

Initial Data Collection and Normalization

More Information Needed

Who are the source language producers?

More Information Needed

Annotations

Annotation process

More Information Needed

Who are the annotators?

More Information Needed

Personal and Sensitive Information

More Information Needed

Considerations for Using the Data

Social Impact of Dataset

More Information Needed

Discussion of Biases

More Information Needed

Other Known Limitations

More Information Needed

Additional Information

Dataset Curators

More Information Needed

Licensing Information

The books have been crawled from smashwords.com, see their terms of service for more information.

A data sheet for this dataset has also been created and published in Addressing "Documentation Debt" in Machine Learning Research: A Retrospective Datasheet for BookCorpus

Citation Information

@InProceedings{Zhu_2015_ICCV,
    title = {Aligning Books and Movies: Towards Story-Like Visual Explanations by Watching Movies and Reading Books},
    author = {Zhu, Yukun and Kiros, Ryan and Zemel, Rich and Salakhutdinov, Ruslan and Urtasun, Raquel and Torralba, Antonio and Fidler, Sanja},
    booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
    month = {December},
    year = {2015}
}

@article{rust-etal-2022-pixel,
  title={Language Modelling with Pixels},
  author={Phillip Rust and Jonas F. Lotz and Emanuele Bugliarello and Elizabeth Salesky and Miryam de Lhoneux and Desmond Elliott},
  journal={arXiv preprint},
  year={2022},
  url={https://arxiv.org/abs/2207.06991}
}

Contact Person

This dataset was added by Phillip Rust.

Github: @xplip

Twitter: @rust_phillip