A new project wants to build open, shared AI training data that anyone can inspect

Key points

• OpenWALDO, launched by CentOS and Rocky Linux founder Gregory Kurtzer, is a shared, openly licensed dataset intended as a public foundation for training AI models.
• The project is funded by CIQ, Kurtzer's AI infrastructure company, and currently holds 167.3 billion reference tokens drawn from government records, academic papers and public-domain literature.
• Frontier AI models typically train on tens of trillions of tokens, meaning OpenWALDO's current dataset is a small fraction of what large models need.
• CIQ argues that hidden training data puts businesses at legal and security risk, because no one can verify what copyrighted or restricted material may be baked into a model.
• No company has publicly confirmed training a model on OpenWALDO yet; CIQ did not respond when asked.

Most AI models are built on secrets. The data used to teach them is almost never disclosed to the people who buy and use those models. A new project called OpenWALDO, first reported by The Register AI, wants to change that.

OpenWALDO stands for Open Weights, Artifacts, Licenses, Data, Origins. The idea is straightforward: build a shared, publicly inspectable set of training data (the raw text AI models learn from), so any company or researcher can use it as a verified starting point.

Why does secret training data matter to ordinary businesses?

Hidden training data is a liability, not just a philosophical problem. If a model learned from copyrighted books or outputs from rival AI systems, the businesses building products on top of it could face legal exposure without knowing it.

CIQ, the AI infrastructure firm behind OpenWALDO, put it plainly: "There is often no way to know what data trained a given model, under what license, or with what consent."

For a hospital buying an AI tool or a law firm using an AI assistant, that opacity is a real risk. You cannot audit what you cannot see.

There's also a waste problem. Every AI lab training its own model in secret repeats work others have already done. A shared, verified dataset, the OpenWALDO team argues, would let companies skip duplicated effort and layer their own private data on top. A lab can take the dataset as a baseline, add proprietary material, and ship a product with a clear, auditable line back to its sources.

The transparency concern isn't new to this beat. Our 7 August story on the SAFE framework found 120 companies trying to solve a related problem: AI security failures that nobody discloses until it's too late. OpenWALDO is attacking the upstream version of that same opacity.

How does OpenWALDO compare to what big AI labs use?

The gap is real. See the numbers:

A token is roughly three-quarters of an English word, so 167 billion tokens is a large library but still a fraction of what models like GPT-4 or Meta's Llama series consumed during training.

Kurtzer draws a direct parallel to Linux, the open-source operating system that now powers most of the internet's servers. Critics once called open-source code insecure. It won anyway.

"Linux didn't win by being certified safe," Kurtzer said. "It won by being inspectable, forkable, and community validated."

Whether OpenWALDO follows that path is genuinely uncertain. Good intentions are common in open-source AI; sustained adoption is not.

If your business depends on an AI tool, ask your vendor what data trained it and whether they can show you the licence. You probably won't get a full answer today. Asking anyway puts you ahead of most buyers, and that question is exactly why projects like this one get built.