bert-large-nordic-pile-1M-steps: benchmark results
AI Sweden's 24-layer BERT-large masked-language encoder, trained on the Swedish subset of the Nordic Pile using a Wafer-Scale Engine. It is an encoder checkpoint for masked-token prediction and downstream adaptation, distinct from the autoregressive GPT-SW3 releases. Provider: AI Sweden. Access: Open.
Aggregate Rating 1478 ± 39, rank #2694 of 10916 rated models, from 27 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| EuroEval Swedish | 76.25 | Average Score (%) | 99.3 |
| EuroEval Swedish NLU | 76.25 | NLU Average Score (%) | 99.3 |
| EuroEval Swedish NLU - SUC3 | 79.99 | Micro F1 excluding MISC (%) | 98 |
| EuroEval Swedish NLU - ScaLA SV | 72.07 | Matthews correlation coefficient (%) | 97.8 |
| EuroEval Norwegian NLU - NorNE NB | 85.74 | Micro F1 excluding MISC (%) | 96.2 |
| EuroEval Faroese NLU - FONE | 82.96 | Micro F1 excluding MISC (%) | 94.8 |
| EuroEval Icelandic NLU - MIM-GOLD NER | 64.07 | Micro F1 excluding MISC (%) | 79.5 |
| EuroEval Danish NLU - Dansk | 62.1 | Micro F1 excluding MISC (%) | 78 |
| EuroEval Swedish NLU - Swerec | 76.69 | Matthews correlation coefficient (%) | 67.7 |
| EuroEval Norwegian | 44.23 | Average Score (%) | 61.3 |
| EuroEval Icelandic | 30.79 | Average Score (%) | 59 |
| EuroEval Danish NLU - ScaLA DA | 34.29 | Matthews correlation coefficient (%) | 56.4 |
Interactive version: theaggregate.ai/model?slug=bert-large-nordic-pile-1m-steps · How It Works · Data refreshed daily, snapshot 2026-11-08.