Vör finds the personal data

Last week I borrowed an Nvidia DGX Spark to test. Its GPU shares system memory with the CPU, with about 122 GiB visible to the operating system. That was enough for LoRA fine-tuning of a 31-billion-parameter model on one desktop. Training and evaluation ran from 23 to 26 October.

Two models came out of it, both built on Google's Gemma 4. Norn v4 is a general Icelandic model that inflects nouns correctly in 85.6% of 480 BÍN test questions, according to my evaluation of the unquantised model. The base model gets 70%. The other model is called Vör and does one thing. It reads Icelandic text and looks for personal data in it.

Vör labels each item with its type and legal basis. A name and a kennitala are general personal data under article 3 of Act no. 90/2018. Health, union membership and religious belief are sensitive. Criminal records and financial data get their own classes. Finally it says whether the document as a whole is sensitive.

On 317 test documents Vör found 98.7% of distinct normalized personal-data text spans (1,428/1,447), without requiring the correct type label and 99.7% of direct identifiers. I also tested names it never saw in training, and it found 324 of 327. This supports generalisation to unseen names within the synthetic test distribution.

The test documents are synthetic, like all the data it learned from, and no real person appears in them. The next step is to check a hundred documents by hand and then test it on real ones. It is also weaker on single sentences than on full documents, and criminal records are the class it finds least often.

Both models are open on Hugging Face and run in Ollama on a machine with a large GPU (24 GB or more).

• Vör on Hugging Face
• Norn v4 on Hugging Face

The name comes from the Prose Edda, where Vör is one of the goddesses. "She is wise and inquisitive, so that nothing can be hidden from her."