The main constraint was licensing rather than regulations like GDPR. To ensure compliance, we restricted ourselves to datasets with permissive licenses.
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.
I would think it's the fact that there's no secret sauce except data and compute to get to a decent model along with a good team. It's a Nemtron 3 Nano architecture as far as I understand, and of course there seems to be some special data curation for German.
