Overview
This checkpoint continues the pre-training of answerdotai/ModernBERT-base on Scandinavian text, extending the model’s knowledge with ~1.2 trillion additional masked-language-model (MLM) tokens drawn from The Nordic Pile and SWEb while preserving the original 8k token context window.
This is a research artefact and is only intended for research purposes.
Our tokenizer is trained from scratch on a subset of 11 985 103 472 tokens.
The training is done in one stage with 8192 tokens per sample for the whole run.
Data Sources
| Corpus | Size | Selected Languages | Highlights |
|---|---|---|---|
| The Nordic Pile | 1.2 TB raw text | sv, no, da, is | Nine diverse categories (CC, Wikipedia, Books, Code, etc.), filtered and deduplicated for high quality |
| SWEb | 1 T+ tokens (~3.6 TB) | sv, no, da, is | 98 Common-Crawl snapshots with model-based HTML extraction; 1.2 B documents |
Training Setup
| Setting | Value |
|---|---|
| Parameters | 150 M |
| Context length | 8 192 tokens (RoPE + local-global attention) |
| Tokens processed | 1.20 × 1012 |
| Tokens per batch | 1 572 864 |
| Global batch | 192 sequences (micro-batch = 3) |
| Optimizer & schedule | Decoupled StableAdamW, lr 2 e-4, cosine decay (1 % warm-up) |
| Precision | AMP-bf16 |
| Hardware | 8 nodes × 8 AMD MI250X GPUs (64 GPUs) on the EuroHPC LUMI-G system |
See training details here
Training Stats
[token=1198511677292/1198510347252]:
Train time/batch: 873585
Train time/sample: 167728320
Train time/batch_in_epoch: 3558
Train time/sample_in_epoch: 683136
Train time/token: 1198510256276
Train time/token_in_epoch: 4882888303
Train trainer/device_train_microbatch_size: 3
Train loss/train/total: 0.9966
Train throughput/batches_per_sec: 1.3117
Train throughput/samples_per_sec: 251.8442
Train throughput/device/batches_per_sec: 0.0205
Train throughput/device/samples_per_sec: 3.9351
Train throughput/tokens_per_sec: 1804244.5198
Train throughput/device/tokens_per_sec: 28191.3206
Train time/train: 184.5555
Train time/val: 0.0000
Train time/total: 184.5555
Train lr-StableAdamW/group0: 0.0000
Train lr-StableAdamW/group1: 0.0000
Intended Use
This is a research artefact and is only intended for research purposes.
- Fill-mask inference, embedding extraction and fine-tuning for Scandinavian downstream NLP tasks (classification, NER, QA, etc.).
- Drop-in replacement for BERT-style encoders (omit
token_type_ids).
Fill-mask
from transformers import pipeline
unmasker = pipeline('fill-mask', model='AI-Sweden-Models/ModernBERT-base')
unmasker("Huvudstaden i Sverige är [MASK].")
[{'score': 0.0629318505525589,
'token': 2961,
'token_str': ' Stockholm',
'sequence': 'Huvudstaden i Sverige är Stockholm.'},
{'score': 0.03635135293006897,
'token': 49763,
'token_str': 'awesome',
'sequence': 'Huvudstaden i Sverige är awesome.'},
{'score': 0.03006783314049244,
'token': 751,
'token_str': ' stor',
'sequence': 'Huvudstaden i Sverige är stor.'},
{'score': 0.029827557504177094,
'token': 71,
'token_str': 'a',
'sequence': 'Huvudstaden i Sverige är a.'},
{'score': 0.019739385694265366,
'token': 79,
'token_str': 'i',
'sequence': 'Huvudstaden i Sverige är i.'}]
Limitations & Biases
- Web corpora can contain noise, stereotypes and sensitive content despite filtering.
- RoPE extrapolation beyond 8 k tokens is untested and may degrade.