NbAiLab/borealis-open-27b
NbAiLab/borealis-open-27b is a 27-billion parameter instruction-tuned language model developed by the National Library of Norway, based on the Gemma 3 architecture. This open-release model is specifically fine-tuned for Norwegian-centric instruction following, excelling in tasks like drafting, summarization, and question answering in Norwegian. It supports a 32K token context length and is designed for general-purpose Norwegian language applications. The model is part of the Borealis family, focusing on Norwegian language understanding and generation.
Loading preview...
Model Overview
NbAiLab/borealis-open-27b is a 27-billion parameter instruction-tuned model from the National Library of Norway (NbAiLab), built upon the Gemma 3 family of models. This is an "open release" variant, meaning it does not include certain copyright-protected material used in the full Borealis models. It is specifically fine-tuned for Norwegian-centric instruction following.
Key Capabilities
- Norwegian Language Focus: Optimized for assistant-style tasks, writing, summarization, and question answering in Norwegian.
- Instruction Following: Designed to follow instructions for various Norwegian language tasks.
- Safety Alignment: Incorporates "prompt baking" and weighted merging of SFT and aligned models to balance quality, usefulness, and safer behavior.
Training and Evaluation
The model uses a supervised fine-tuning (SFT) dataset, NbAiLab/aurora-sft-open, prepared by the National Library of Norway. Evaluation is conducted using NorEval, MMLU-English, and NbAiLab's proprietary nb-gpt-bench suite.
Intended Use Cases
- Norwegian-centric assistant tasks (drafting, summarization, Q&A).
- Assessment and improvement of Norwegian writing style.
- Evaluation of language coverage for Norwegian, Bokmål, and Nynorsk.
Limitations
Users should be aware that the model may hallucinate, produce incorrect information, or generate biased/inappropriate outputs. Performance outside Norwegian and English use cases is not fully characterized, and it is not intended as a reasoning model.