The one exception was the Genios dataset (a high-quality German corpus), which we used during pretraining under a separate licensing agreement. Beyond that, the growing availability of high-quality permissive datasets meant we could assemble sufficient training data without compromising on quality.