OpenWALDO Brings Open Source Discipline to AI Training Data
AI Training Data Needs a Trust Layer
OpenWALDO puts a spotlight on one of the least transparent parts of AI development. Even open-weight models can leave users guessing about what data shaped the model, what licenses apply and whether unwanted behavior was introduced during training. Gregory Kurtzer, Founder and CEO of CIQ, explains why that opacity creates risk for enterprises, developers and model builders.
The project applies open source principles to AI training data. OpenWALDO is designed to make training inputs named, reviewable, attributable and verifiable. That matters because model behavior depends on more than weights. It also depends on the data, artifacts, licenses and origins that shaped the model before it reached users.
An AI Bill of Materials Creates Accountability
OpenWALDO also introduces the idea of an AI bill of materials. That record can connect a model back to its sources, licenses, training data and supporting artifacts. For organizations trying to understand model risk, that traceability can become a critical control.
Kurtzer compares the approach to the transparency that made open source software valuable. Developers can inspect source code and understand what a program is built to do. AI models do not offer the same visibility today. A shared corpus with review processes, licensing evidence and provenance can help close that gap.
Open Data Can Reduce Redundant Work
The conversation also explores the cost of repeatedly training models on similar data. Every AI lab and enterprise team does not need to rebuild the same foundation from scratch. A community-managed corpus can give model builders a trusted baseline, while still allowing companies to add proprietary or domain-specific data where needed.
OpenWALDO uses Git for the index and relies on cryptographic validation for larger data objects. Contributors can host data in their own storage buckets, while the index records what the data is and where it came from. Developer Certificate of Origin signoffs add another layer of accountability.
Smaller Models Need Better Foundations
OpenWALDO may also help teams that want to build smaller, more focused models. Kurtzer notes that specialized models can be more practical than massive general-purpose models for many enterprise use cases. Still, building models requires skill, tooling, compute and community knowledge.
That is where a shared open source project can lower the barrier. OpenWALDO gives researchers, companies and practitioners a place to contribute, learn and improve the training-data foundation together. It also gives companies a new way to ensure their public product documentation and knowledge can be represented accurately in future AI systems.
