Vör and Norn
Icelandic language models, trained on my own machine and a borrowed DGX Spark
Norn inflects Icelandic correctly and Vör finds personal data in Icelandic documents. Both models are open on Hugging Face.
Norn is a series of Icelandic language models I train on Hercules, my own machine, and on a DGX Spark I borrowed for testing. The fourth version is built on Gemma 4 with 31 billion parameters and scores 85.6% on 480 noun forms drawn from the Database of Icelandic Morphology, according to the model card for the unquantized checkpoint. The base model gets 70%.
Vör is its specialised sibling. It reads Icelandic documents and labels detected personal-data spans under Act no. 90/2018 on data protection, whether a name, a kennitala, a health condition or a criminal record. On 317 synthetic test documents, it recovered 1,428 of 1,447 distinct normalized gold text spans (98.7% recall). This text-matching measure does not require the correct entity type and is not overall accuracy. The test documents are synthetic, and real documents are the next test.
Both models are open on Hugging Face: Vör and Norn v4. More in the blog post.
Technology Stack
Related Projects
View all projects →
BORG
The University of Akureyri's hybrid AI platform, combining commercial models with on-premise infrastructure with an option for processing on the university network.
Smarason.is
A personal site moved home: Next.js 16 + Payload on its own Hetzner edge with Caddy, no cloud content service, no first-party analytics. The architecture is the philosophy — sovereignty over your own data.
Institutional Knowledge Graph
Small organisations run on invisible architecture — governance, systems, policies, dependencies that live in people’s heads. This is the complete methodology for making that architecture queryable: Neo4j, JSON-LD, SHACL, and a pipeline that is always safe to re-run.