PD12M (Public Domain 12M)

PD12M is the largest image-text dataset drawn only from material that carries a Public Domain Mark or a Creative Commons Zero (CC0) waiver, meaning the source images come with no copyright strings attached. It holds 12.4 million image-caption pairs, which puts it on a par with widely used web-scraped sets like Conceptual Captions, but without the licensing risk of scraping images under unknown terms. A curated 3.3 million item subset, PD3M, offers the same clean provenance at a more manageable size.

The dataset lives on Hugging Face and is distributed as metadata: parquet files listing image URLs, captions, dimensions, and licence tags rather than the pixels themselves. Crucially, the images are rehosted on dedicated cloud storage separate from their original websites, which tackles link rot, the slow decay of web-scraped datasets as the pages they point to vanish. To use it you stream the metadata, then fetch the images from those stable URLs, so budget storage and download time for the image payload.

For a RAG project, PD12M suits multimodal retrieval: building a searchable index of images by their captions and embeddings, powering visual search, or supplying a clean, commercially safe corpus for training and fine-tuning text-to-image and image-understanding models. It is the standout choice whenever copyright cleanliness is a hard requirement.

Watch a few things. The captions describe the images at varying depth and some are machine-generated, so quality is uneven and worth spot-checking for your use. Public domain material also skews towards older artwork, museum holdings, and historical photography, which is wonderful for some domains and a poor fit for contemporary or product imagery. On licensing you are on unusually firm ground: the compilation is released under CDLA-Permissive-2.0 and the underlying images are public domain or CC0, so commercial use is far clearer here than with scraped sets like LAION.