OpenWALDO launches shared AI training data corpus with provenance records
CIQ-backed OpenWALDO wants to make AI training inputs traceable, but its 167.3 billion-token corpus remains early-stage.
By Dominic Okoye · Staff Writer
· 3 min read
OpenWALDO, a CIQ-sponsored project founded by Rocky Linux and CentOS creator Gregory Kurtzer, has launched a shared corpus for AI training data that is intended to record provenance and licensing information alongside the material. The OpenWALDO AI training data effort is trying to address a layer that most downloadable models do not expose: the documents, rights assertions and training process behind their weights.
CIQ says WALDO stands for Weights, Artifacts, Licenses, Data and Origins. The project proposes a community-governed corpus to which researchers, companies and individuals can contribute, rather than having model teams separately collect, clean, license and maintain overlapping foundational material.
CIQ is funding the effort, and Kurtzer is leading it. The company has not presented OpenWALDO as a frontier-model project at launch. It is building the data corpus and associated tooling for other labs and developers to use in training models.
How is OpenWALDO different from open-weight AI models?
Open-weight models make a model’s learned parameters available for download and use. That does not necessarily disclose the code, methods, recipes or data needed to reproduce the model. OpenWALDO’s stated standard is broader: an inspectable supply chain for the underlying inputs and training process, not weights alone.
Under CIQ’s proposed workflow, the project’s index will record where a training item came from, the rights asserted over it, its contributor and the exact bytes included in a version. A resulting AI Bill of Materials could connect a trained model to its source materials, licenses and training runs. CIQ says the public index is maintained in Git; larger objects sit in federated storage and are checked with content hashes; and contributors sign off using the Developer Certificate of Origin.
Those mechanisms describe the project’s design, rather than an independently audited compliance system. They could give a model builder a documented baseline, but they do not independently establish that every rights claim, consent record or data-quality decision is correct.
How large is OpenWALDO’s training corpus?
The Register reported that the corpus contains 167.3 billion reference tokens drawn from government records, open-source academic papers, mailing lists and public-domain literature. That is substantial as a public collection, but it remains far smaller than the tens of trillions of tokens the publication said are used for frontier-model and open-weight training.
CIQ says a lab or company could take the corpus and its bill of materials, add proprietary data, and retain a traceable record of the inputs it used. The value proposition is a reusable, reviewable base rather than a claim that the initial corpus can support a competitive frontier model.
The central execution test is adoption. Available reporting has not identified a trained model or trained-model result built from the corpus; The Register said CIQ did not answer its question about whether one had been trained. OpenWALDO’s usefulness will depend on whether contributors expand the collection, rights and provenance records hold up to scrutiny, and model builders decide the documentation is worth incorporating into their pipelines.
This story draws on original reporting from The Register.
