The org behind the data set used to train Stable Diffusion claims it has removed CSAM

The German nonprofit LAION claims to have cleaned its data set used for training AI models of CSAM links. This announcement follows recommendations from multiple watchdog groups. The updated data set, Re-LAION-5B, aims to ensure safer research.

Background on LAION

LAION, a German research organization, is behind the data set used to train Stable Diffusion and other AI models. They recently released Re-LAION-5B, a cleaned version of their initial data set, LAION-5B.

This cleanup follows guidelines from the nonprofit Internet Watch Foundation, Human Rights Watch, and the Canadian Center for Child Protection. The new data set is designed for safer, more ethical AI model training.

Details of Re-LAION-5B

The new Re-LAION-5B data set is a revision, including about 5.5 billion text-image pairs. It was released under an Apache 2.0 license.

Two versions are available: Re-LAION-5B Research and Re-LAION-5B Research-Safe, the latter also removing additional NSFW content. Both versions were filtered for thousands of CSAM links.

Investigations and Findings

In December 2023, Stanford Internet Observatory found that the initial LAION-5B data set contained 1,679 links to illegal images.

These links came from social media posts and adult websites and included inappropriate content like pornographic imagery and harmful stereotypes.

The presence of such content led to recommendations for model deprecation and distribution cessation.

Community and Industry Response

Following the Stanford report, LAION took the initial data set offline temporarily. They assured that their data sets were for research purposes only.

However, companies like Google and Stability AI have used LAION data sets for commercial purposes. This highlights the challenges in controlling data set usage.

Steps Taken for Data Cleaning

LAION worked with multiple partners to clean their data set. They matched suspected CSAM links with lists from watchdog groups.

In total, 2,236 links were removed, including 1,008 links identified by the Stanford report.

LAION recommends all users of the old LAION-5B data set to switch to the new Re-LAION-5B for better safety.

Impact on AI Training

The cleaned data set aims to provide a safer and more ethical foundation for AI model training.

However, the issue of inappropriate content in data sets remains a challenge for the AI community.

The efforts to clean the data set show a commitment to responsible AI research and development.

Conclusion and Future Steps

LAION’s efforts to clean their data set highlight the ongoing need for ethical AI data usage.

The release of Re-LAION-5B is a step forward, but the industry must remain vigilant.

Future AI model training will benefit from these improved data sets.

The release of Re-LAION-5B is a crucial step in ensuring safer AI model training. However, continuous efforts are essential to maintain ethical standards. The industry must adopt and follow these guidelines closely for AI advancements.

About The Author

Emmanuel Kesse

See author's posts