OpenWALDO launches transparent AI training dataset with 167B tokens

Gregory Kurtzer, the founder of CentOS and Rocky Linux, has announced OpenWALDO, a new open-source AI training dataset initiative funded by his company CIQ, which also sponsors Rocky Linux. The project name stands for Open Weights, Artifacts, Licenses, Data, Origins, and reflects Kurtzer's effort to apply open-source principles to AI model development—an area he argues has yet to embrace true openness despite the recent prominence of open-weight models.

The core problem OpenWALDO addresses is that most downloadable open-weight models, while their code is publicly available, rely on closed-source training data. This data often combines copyrighted material, distilled responses from other proprietary models, and user-generated content obtained without clear consent or license disclosure. CIQ argues this opacity creates multiple risks: customers have no way to audit what trained their model or under what licenses, hidden data sources can taint model behavior and put software stacks at risk, and the industry wastes computing resources by each organization separately gathering and processing training data. A single shared public baseline would eliminate this duplication and allow improvements to benefit all future models trained on it.

Kurtzer positions OpenWALDO as a solution that mirrors the success of open-source software. Companies and labs can take the OpenWALDO corpus as a verified baseline, add their own proprietary data, build their model, and maintain a clear auditable trail back to the source of every component. This approach addresses concerns about security and misuse that frontier labs have raised about open-weight models; Kurtzer counters that open-source software faced identical skepticism but won through inspectability, forkability, and community validation—not safety claims. "Linux didn't win by being certified safe. It won by being inspectable, forkable, and community validated," he said. "AI is missing that same property, and OpenWALDO is how we build it."

The timing reflects industry momentum behind open-weight models, particularly from China, which are closing in on the capabilities of closed-source frontier models like ChatGPT and Claude. Many businesses now question why they should pay for AI services they don't own, cannot truly control, and have no visibility into. The OpenWALDO dataset currently contains 167.3 billion reference tokens gathered from government records, open-source academic papers, mailing lists, and public domain literature—a substantial amount, but a fraction of the tens of trillions used to train frontier models. CIQ did not confirm whether any organization has yet trained a model on the dataset. Adoption remains uncertain; while the project reflects legitimate industry pain points, persuading competitors to share training work represents a major cultural shift for an industry accustomed to treating data as a proprietary advantage.