You can now train AI search agents without paying for expensive APIs — a dataset proves it works

Researchers built a 20,000-question dataset for training AI search agents without using paid API services, proving that synthetic data generated cheaply can work as well as expensive human-annotated datasets. This means smaller labs and companies can now train their own search-capable AI models instead of relying on expensive proprietary systems.

Training data for AI search agents has been locked behind expensive human annotation or proprietary APIs — making it impossible for anyone without deep pockets to build competitive systems. This dataset shows that you can generate training data for free using a pipeline of automated reasoning, self-checking, and web verification, and the resulting trained models perform competitively. That opens the door to smaller organizations building their own search-capable AI instead of buying access from the handful of companies that can afford annotation at scale.