NbAiLab/borealis-open-270m
NbAiLab/borealis-open-270m is a 270 million parameter instruction-tuned model from the Borealis family, developed by the National Library of Norway. Based on the Gemma 3 architecture, this open release model is fine-tuned for Norwegian-centric instruction following, assistant-style tasks, and writing assessment. It specializes in tasks like drafting, summarization, and Q&A in Norwegian, including Bokmål and Nynorsk.
Loading preview...
What is Borealis Open 270M?
NbAiLab/borealis-open-270m is a 270 million parameter instruction-tuned model developed by the National Library of Norway (NbAiLab). It is part of the Borealis family of models and is an "open release" variant, meaning it does not include material from specific rights-holder agreements. This model is built upon the Gemma 3 architecture, specifically fine-tuned from google/gemma-3-270m-it.
Key Capabilities
- Norwegian-centric instruction following: Optimized for tasks requiring understanding and generation in Norwegian.
- Assistant-style tasks: Proficient in drafting, summarization, and question answering.
- Writing assessment: Can be used for assessing and improving Norwegian writing style and quality.
- Multilingual focus: Evaluates behavior and language coverage for Norwegian, Bokmål, and Nynorsk.
Training and Alignment
The model underwent Supervised Fine-Tuning (SFT) using instruction data prepared by the National Library of Norway, focusing on Norwegian-centric assistant behavior. Safety alignment is achieved through "prompt baking" and weighted merging of SFT and aligned models, aiming to balance quality, usefulness, and safer behavior. The SFT dataset used is NbAiLab/aurora-sft-open.
When to Use This Model
This model is particularly suitable for applications requiring a compact, efficient model with strong capabilities in Norwegian language processing. It is intended for:
- Norwegian-centric assistant tasks.
- Assessment and improvement of Norwegian writing.
- Research into Norwegian language coverage and behavior.
Users should be aware of general generative model limitations, including potential for incorrect or biased outputs, and should not use it for safety-critical applications without additional safeguards.