File size: 4,326 Bytes
2fa40b5 |
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 |
---
license: mit
task_categories:
- fill-mask
tags:
- pretraining
- encoder
- multilingual
---
# mmBERT Mid-training Data
[](https://opensource.org/licenses/MIT)
[](https://arxiv.org/abs/2509.06888)
[](https://huggingface.co/collections/jhu-clsp/mmbert-a-modern-multilingual-encoder-68b725831d7c6e3acc435ed4)
[](https://github.com/jhu-clsp/mmBERT)
> **Phase 2 of 3**: High-quality mid-training data mixture (600B tokens) with context extension to 8192 tokens.
This dataset contains the mid-training phase data used to train all [mmBERT encoder models](https://huggingface.co/collections/jhu-clsp/mmbert-a-modern-multilingual-encoder-68b725831d7c6e3acc435ed4). This phase focuses on higher quality data sources and extends the context length from 1024 to 8192 tokens. The data is provided in **MDS format** ready for use with [Composer](https://github.com/mosaicml/composer) and the [ModernBERT training repository](https://github.com/answerdotai/ModernBERT).
## ๐ Data Composition
| Data Source | Tokens (B) | Percentage | Description |
|:------------|:-----------|:-----------|:------------|
| FineWeb2 | 506.7 | 84.3% | High-quality multilingual web crawl data |
| DCLM (Dolmino) | 40.0 | 6.7% | Filtered high-quality English web data |
| Starcoder | 17.2 | 2.9% | Code repositories and files |
| Arxiv | 5.4 | 0.9% | Academic preprints |
| Dolmino Math | 4.3 | 0.7% | Mathematical content |
| Books | 3.9 | 0.7% | Literature and reference books |
| PeS2o | 3.2 | 0.5% | Scientific papers |
| Tulu Flan | 3.1 | 0.5% | Instruction-following data |
| StackExchange | 3.0 | 0.5% | Q&A forums |
| StackExchange (Dolmino) | 2.8 | 0.5% | Curated Q&A content |
| Wikipedia (MegaWika) | 1.2 | 0.2% | Encyclopedia articles |
| **Total** | **600.8** | **100.0%** | High-quality data for context extension |
## ๐ Language Coverage
This phase covers **110 languages** plus code, with inverse temperature sampling at ฯ=0.5. Expands from the initial 60 languages to include:
- **Additional mid-resource languages**: Uzbek, Bosnian, Catalan, Albanian, and 46 others
- **Enhanced quality**: Uses filtered FineWeb2-HQ and higher quality DCLM
- **Longer contexts**: Optimized for 8192 token sequences
## โ๏ธ Key Features
- **Context Extension**: RoPE base frequency adjusted to 160k for 8192 token support
- **Quality Upgrade**: Switches to filtered, higher-quality versions of datasets
- **Reduced Masking**: Mask rate lowered to 15% (from 30% in pre-training)
- **Language Expansion**: Adds 50 new languages while maintaining data quality
## ๐ Usage
For mid-training, see the ModernBERT repo: https://github.com/AnswerDotAI/ModernBERT
### Direct Access
```python
from streaming import StreamingDataset
# Load the streaming dataset
dataset = StreamingDataset(
remote='https://huggingface.co/datasets/jhu-clsp/mmbert-midtraining',
local='/tmp/mmbert-midtraining-data',
shuffle=True
)
# Access samples
for sample in dataset:
text = sample['text']
# Process your data...
```
## ๐ Related Resources
- **Models**: [mmBERT Model Suite](https://huggingface.co/collections/jhu-clsp/mmbert-a-modern-multilingual-encoder-68b725831d7c6e3acc435ed4)
- **Phase 1**: [Pre-training Data](https://huggingface.co/datasets/jhu-clsp/mmbert-pretrain-p1-fineweb2-langs) (2.3T tokens)
- **Phase 3**: [Decay Phase Data](https://huggingface.co/datasets/jhu-clsp/mmbert-decay) (100B tokens)
- **Checkpoints**: [Training Checkpoints](https://huggingface.co/datasets/jhu-clsp/mmbert-checkpoints)
- **Paper**: [Arxiv link](https://arxiv.org/abs/2509.06888)
- **Code**: [GitHub Repository](https://github.com/jhu-clsp/mmBERT)
## Citation
```bibtex
@misc{marone2025mmbertmodernmultilingualencoder,
title={mmBERT: A Modern Multilingual Encoder with Annealed Language Learning},
author={Marc Marone and Orion Weller and William Fleshman and Eugene Yang and Dawn Lawrie and Benjamin Van Durme},
year={2025},
eprint={2509.06888},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2509.06888},
}
``` |