Model Showcase
Norwegian Legal AI — fine-tuned on 7,603 enhanced Q&A pairs using knowledge distillation from a 24B teacher model. Built for precision, grounded in real law.
Training Pipeline
Five stages transform 2,847 legal documents into a specialised Norwegian law model.
14 scrapers across Norwegian courts, EU law, and regulatory bodies
2,847 documents
Automated question-answer pair extraction from legal corpus
7,603 pairs
NorwAI-24B teacher model rewrites and enriches every answer
Knowledge Distillation
4-bit quantised LoRA training on RTX PRO 6000 (96 GB VRAM)
31,842 train examples
Merge LoRA, convert to GGUF Q5_K_M via llama.cpp
5.1 GB model
Knowledge Distillation
The NorwAI 24B teacher model rewrites every Q&A pair with richer legal context, precise terminology, and professional Norwegian language. The student model learns to produce this quality at a fraction of the size and cost.
The result: a 5.1 GB model that runs on consumer hardware yet answers Norwegian legal questions with the depth of a much larger system.
Training Metrics
Head-to-Head
Same questions, two models. Real outputs — unedited. Click each topic to see the difference.
"Hva er barnets beste-prinsippet i norsk barnevernrett, og hvordan anvendes det i praksis?"
Evaluation
| Criterion | Base Qwen 7B | dobetter-norge-v2 |
|---|---|---|
| Legal Citations | None | HR-1986-573-A, § 4-1, § 4-2, § 5-10, FN, ECHR, Bufdir |
| Norwegian Quality | Poor — fake words, Danish mix | Professional legal register |
| Depth & Detail | Generic bullet-point overview | Multi-paragraph legal analysis |
| Practical Guidance | Vague, no concrete examples | Omsorgsovertakelse, ekspertvurdering, Bufdir-metode |
| Factual Accuracy | Confuses adults and children; wrong law refs | Cites real law chapters and case numbers |
| Overall | D | A |
Get The Model
Qwen 2.5 7B · Q5_K_M · 5.1 GB · Norwegian Legal AI
dobetter-norge-v2 powers the CorpusAI platform. Ask it anything about Norwegian law.