OpenWALDO launches transparent AI training dataset with 167B tokens

OpenWALDO addresses a growing tension in the AI industry: despite the rise of downloadable open-weight models, their training data remains opaque. Companies and labs typically draw from copyrighted material, distilled responses from other models, or user-generated content obtained without clear consent—sources that are neither disclosed nor auditable. This opacity creates two problems the project aims to solve. First, it poses practical risk: customers cannot know whether a model's foundation is tainted, putting their software stacks in jeopardy. Second, it wastes resources: without a shared baseline, every organization duplicates the effort of curating and validating training data.

Kurtzer's approach borrows from open-source software's playbook. Rather than keeping the dataset proprietary or claiming safety through obscurity, OpenWALDO makes sources traceable and forkable. Companies can then layer their own proprietary data on top of the public baseline, creating a clear audit trail from final model back to its foundations. This model has already resonated with businesses questioning why they pay for closed AI services they cannot control or inspect, especially as open-weight models from China begin matching the capabilities of proprietary alternatives like ChatGPT and Claude.

However, OpenWALDO faces a steep climb. The 167.3 billion tokens currently available are a drop next to the tens of trillions used by frontier labs, and it remains unclear whether any model has been trained on the dataset yet. Adoption requires the AI industry to abandon its closed-data practices—a shift that took decades for open-source software and may prove harder in an AI landscape where companies view proprietary training data as a core competitive advantage.