LAION-5B taken offline over CSAM links and re-released as Re-LAION-5B

Background

LAION-5B is a large open dataset that has been used to train image models. It is not a collection of pictures. LAION's datasets hold indexes: lists of links to images together with the alt text that came with them, curated from Common Crawl^[1]. What sits in the dataset is therefore a link, not an image file^[1].

The December 2023 investigation

In December 2023 the Stanford Internet Observatory published an investigation of LAION-5B. It found that the dataset — specifically the 400-million-pair subset LAION-5B 400M — included at least 1,679 links to illegal images, scraped from social media posts and popular adult websites^[1]. According to AP reporting, the report prompted LAION to remove the dataset immediately^[2]. LAION said it would take LAION-5B offline temporarily^[1].

The report went beyond the dataset itself. It recommended that models trained on LAION-5B should be deprecated and that their distribution should cease where feasible^[1]. The reasoning follows from how training works: once a model has learned from a dataset, repairing the dataset afterwards does not repair the model.

Re-LAION-5B

On August 30, 2024, LAION released Re-LAION-5B, which it described as a re-release of LAION-5B that had been thoroughly cleaned of known links to suspected child sexual abuse material (CSAM)^[1][2]. In all, 2,236 links to suspected CSAM were removed after matching with link and image hash lists provided by partners, and those removals subsume the 1,008 links found in the Stanford Internet Observatory report^[1].

Re-LAION-5B contains around 5.5 billion text-image pairs and was released under an Apache 2.0 license^[1].

Why it matters

The sequence was quick by the standards of dataset maintenance: a report in December 2023, a takedown, and a cleaned re-release in August 2024^[1][2].

The episode put the assembly and screening of web-scale training data in public view. A dataset built by crawling is only as clean as the filtering applied to it afterwards, and at this size that filtering is done by matching links and images against hash lists supplied by partners^[1]. The number removed was larger than the figure listed in the Stanford report and covered it^[1], which indicates that the cleanup drew on more than that single investigation.

The case also shows that a dataset is one link in a chain: an external report, a takedown, a cleaned re-release, and advice aimed at the people who trained on the original. Under the Stanford recommendation, models trained on the earlier release were candidates for deprecation^[1]. For users of an open dataset, that means a correction at the data layer can turn into work at the model layer.

For open datasets in general, the episode points at an awkward property of link indexes. The index itself is a list of links^[1], while the material it points to lives elsewhere and may be illegal. Re-releasing a cleaned version under a permissive license^[1] is one way to keep such a resource in use.