Rocky Linux founder launches OpenWALDO to bring transparency to AI training data

Gregory Kurtzer, the founder of Rocky Linux and co-founder of CentOS, has launched OpenWALDO, a community-governed open-source initiative aimed at bringing transparency and auditability to AI training data. The project's full name—Open Weights, Artifacts, Licenses, Data and Origins—reflects its mission to make training data behave like open-source software dependencies: named, reviewable, versioned, attributable, and verifiable. Backed by Kurtzer's enterprise infrastructure company CIQ, OpenWALDO addresses what Kurtzer calls a critical gap in the AI industry, where training data has remained largely closed and unexamined even as open-weights models and open-source AI code have become more prevalent.

The technical approach centers on three mechanisms: Git-based review for metadata, content-addressed storage for underlying data objects, and an AI Bill of Materials (ABOM) that tracks resolved data and model lineage through training and release. Rather than building its own foundation model, OpenWALDO provides a common base layer—a licensed, provenance-tracked training corpus that model builders can use, extend with proprietary data, and carry into their own training runs. This structure is designed to let organizations focus resources on differentiation rather than repeatedly assembling basic training datasets. At launch, OpenWALDO's public corpus indexes 124 billion reference tokens across 75.1 million documents, organized into 20 corpora and 1,051 shards, with 18 asserted license identifiers.

The timing reflects growing concern about training data provenance and licensing. The MIT-led Data Provenance Initiative audited more than 1,800 text datasets and found that license information was often omitted or miscategorized, with license-omission rates above 70% and error rates greater than 50%. OpenWALDO also addresses a quality problem: recursive training on outputs from other AI systems, known as model collapse, degrades model performance. By maintaining a public record of training sources, developers can distinguish and select training material more carefully. The project adopts practices familiar from software development—public review, attributable contributions, Developer Certificate of Origin sign-offs, and preserved provenance from data ingestion through release—and will accept contributors from individuals, researchers, institutions, and companies, while CIQ provides commercial support rather than control.

Kurtzer framed the initiative in terms of open-source philosophy: "I've spent my career watching open source turn users into builders, competitors into collaborators, and shared problems into common infrastructure that operates at massive scale." He argued that "Linux didn't win by being certified safe. It won by being inspectable, forkable, and community validated. AI is missing that same property, and OpenWALDO is how we build it." Success will depend on OpenWALDO's ability to attract sustained supplies of high-quality, legally usable data; establish trusted review and governance processes; and persuade AI developers that its provenance records provide enough practical value to become standard in model-development workflows.