Datasets:
language:
- en
pretty_name: PD12M
license: cdla-permissive-2.0
tags:
- image
PD12M
Summary
At 12.4 million image-caption pairs, PD12M is the largest public domain image-text dataset to date, with sufficient size to train foundation models while minimizing copyright concerns. Through the Source.Plus platform, we also introduce novel, community-driven dataset governance mechanisms that reduce harm and support reproducibility over time.
Jordan Meyer Nicholas Padgett Cullen Miller Laura Exline
About
PD12M was built and curated with Source.Plus with the aim of resolving many of the data quality issues that arise in web-scraped training data: the presence of copyrighted material, low quality images and captions, violent or nsfw content, PII, decaying dataset quality via broken links, etc.
PD12M consists of entirely public domain and CC0 licensed images, with automated recaptioning of image data, and quality and safety filtering. Images in PD12M are also hosted on dedicated cloud storage, separate from the original image hosts, to avoid placing an undue burden on those hosts or impacting services for regular users. This also ensures the dataset remains wholly intact over its lifetime.
Overview
This dataset has two components. The first is the metadata, which contains the image urls, captions, image dimensions, etc. The second component are the images.
Metadata
The metadata is made available through a series of parquet files with the following schema:
id: A unique identifier for the image.url: The URL of the image.caption: A caption for the image.width: The width of the image in pixels.height: The height of the image in pixels.mime_type: The MIME type of the image file.hash: The MD5 hash of the image file.license: The URL of the image license.
Additionally, CLIP Vit-L/14 embeddings are provided in the embeddings directory.
Images
The image files are all hosted in the AWS S3 bucket pd12m. The URLs to the images files are all maintained in the metadata files.
Tutorials
License
The dataset is licensed under the CDLA-Permissive-2.0.
Reporting Issues
We've gone through great lengths to ensure the dataset is free from objectionable and infringing content. If you find any issues or have any concerns, please flag the item in Source.Plus, where our review process will remove the infringing material, and find a suitable replacement.