File size: 4,124 Bytes
797d765 1ababe2 797d765 625922a 797d765 74af928 797d765 0bf7382 797d765 0bf7382 797d765 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 |
---
library_name: transformers
license: apache-2.0
language:
- sv
- 'no'
- da
- is
tags:
- masked-lm
- fill-mask
- long-context
- modernbert
pipeline_tag: fill-mask
inference: false
base_model:
- answerdotai/ModernBERT-base
---
## Overview
This checkpoint continues the pre-training of [answerdotai/ModernBERT-base](https://huggingface.co/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](https://arxiv.org/pdf/2303.17183) and [SWEb](https://arxiv.org/pdf/2410.04456) 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 × 10<sup>12</sup> |
| 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](https://github.com/timpal0l/ModernBERT/blob/main/training/trainer_lumi.yaml)
## Training Stats
```python
[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
```python
from transformers import pipeline
unmasker = pipeline('fill-mask', model='AI-Sweden-Models/ModernBERT-base')
unmasker("Huvudstaden i Sverige är [MASK].")
```
```python
[{'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.
## Code to reproduce
* [Training](https://github.com/timpal0l/ModernBERT/tree/main/training)
* [Data Processing](https://github.com/timpal0l/ModernBERT/tree/main/tokenizer) |