notesum.ai
Published at December 9Small Languages, Big Models: A Study of Continual Training on Languages of Norway
cs.CL
Released Date: December 9, 2024
Authors: David Samuel1, Vladislav Mikhailov, Erik Velldal, Lilja Øvrelid, Lucas Georges Gabriel Charpentier, Andrey Kutuzov
Aff.: 1University of Oslo, Language Technology Group

| Benchmark | Language |
NorMistral-11B |
NorwAI-Mistral-7B |
NorwAI-Llama2-7B |
normistral-7b-warm |
NorGPT-3B |
Viking-7B |
Viking-13B |
Mistral-Nemo-12B |
|---|---|---|---|---|---|---|---|---|---|
| Reading comprehension | |||||||||
| Belebele (0-shot) | Bokmål | 56.7 | 33.4 | 38.0 | 37.4 | 26.8 | 27.6 | 28.2 | 62.8 |
| NorQuAD (1-shot) | Bokmål | 76.7 | 63.0 | 39.2 | 64.8 | 3.0 | 48.4 | 57.1 | 76.5 |
| Sentiment analysis | |||||||||
| NoReC (sentence-level; 16-shot) | Bokmål | 90.5 | 88.6 | 86.0 | 84.9 | 49.7 | 77.9 | 79.2 | 86.9 |
| NoReC (document-level; 1-shot) | Bokmål | 91.2 | 81.2 | 79.2 | 82.9 | 51.5 | 80.4 | 86.8 | 89.2 |
| Commonsense reasoning | |||||||||
| NorCommonsenseQA (0-shot) | Bokmål | 59.2 | 52.4 | 48.1 | 48.9 | 35.7 | 43.7 | 49.7 | 44.0 |
| NorCommonsenseQA (0-shot) | Nynorsk | 51.6 | 43.2 | 37.9 | 43.2 | 29.5 | 39.0 | 40.0 | 33.7 |
| World knowledge | |||||||||
| NRK-Quiz-QA (0-shot) | Bokmål | 63.7 | 55.2 | 52.3 | 57.9 | 33.1 | 44.2 | 51.0 | 47.4 |
| NRK-Quiz-QA (0-shot) | Nynorsk | 71.9 | 65.2 | 64.3 | 65.9 | 37.3 | 51.1 | 54.8 | 47.2 |
| NorOpenBookQA (16-shot) | Bokmål | 68.2 | 39.3 | 38.7 | 37.7 | 30.2 | 35.7 | 37.3 | 74.4 |
| NorOpenBookQA (16-shot) | Nynorsk | 60.0 | 36.7 | 34.4 | 33.3 | 38.9 | 22.2 | 26.6 | 73.3 |
| Summarization | |||||||||
| NorSumm (0-shot) | Bokmål | 40.0 | 12.1 | 4.8 | 17.4 | 28.6 | 30.6 | 33.8 | 39.5 |
| NorSumm (0-shot) | Nynorsk | 30.6 | 10.6 | 3.1 | 9.9 | 22.4 | 25.8 | 27.4 | 29.4 |
| Grammatical error correction | |||||||||
| ASK-GEC (16-shot) | Bokmål | 52.6 | 53.2 | 51.4 | 48.7 | 1.8 | 51.1 | 52.4 | 43.9 |
| Language identification | |||||||||
| Scandinavian LID (16-shot) | 98.6 | 96.2 | 93.4 | 77.5 | 40.3 | 68.9 | 65.6 | 59.5 | |
| Translation | |||||||||
| Tatoeba (from English; 16-shot) | Bokmål | 58.8 | 58.7 | 57.9 | 57.2 | 1.8 | 59.7 | 60.0 | 49.6 |
| Tatoeba (from English; 16-shot) | Nynorsk | 48.0 | 47.4 | 47.4 | 44.7 | 2.6 | 45.6 | 45.6 | 35.7 |
| Tatoeba (from English; 16-shot) | Northern Sámi | 45.5 | 24.8 | 26.6 | 14.9 | 0.0 | 4.6 | 9.8 | 3.4 |