NorOLMo 1.0 is a base (not instruction-tuned) large language model, continually pre-trained on Norwegian and Scandinavian data starting from the English OLMo2-13B model.

It is the only modern fully open large language model specifically adapted for Norwegian and Sámi. At the same time, its performance is comparable to other multilingual open models (see the Evaluation section below). "Fully open" here means that the entirety of the trainining data and training pipeline is transparent and reproducible, not only that the model weights are available.

The model was trained for 33 000 steps on around 275 billion tokens. Maximum sequence length was set to 4 096 in the beginning of training, and then was extended to 16 384 starting from the checkpoint step_31000. Intermediate checkpoints are published here as branches. The main branch contains the model's weights after step 33 000 (stage 3).

In a few months, we hope to follow with post-trained versions of NorOLMo.

Below is a comparison of fully-open models supporting Norwegian. The figure shows the aggregate score across all 35 NorEval 1.1 tasks (5 categories, category average). Scores are first averaged within each task category, then averaged across categories. This gives equal weight to each category regardless of how many tasks it contains. Each task score is normalized to a 0–100 scale where 0 = random baseline performance and 100 = perfect score, then averaged across tasks. This accounts for different chance levels across tasks (e.g. 25% for 4-choice QA vs. 50% for binary classification).

More detailed evaluation is evailable in our interactive NorEval dashboard: https://ltgoslo.github.io/llm-dashboard.

Furthermore, an interactive per-checkpoint evaluation with additional ablation studies is available here.

Data ("pretraining data")

Data ("midtraining data")

Training was conducted as a part of the HPLT project.