Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,112 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
task_categories:
|
4 |
+
- fill-mask
|
5 |
+
tags:
|
6 |
+
- pretraining
|
7 |
+
- encoder
|
8 |
+
- multilingual
|
9 |
+
---
|
10 |
+
|
11 |
+
# MMBERT Decay Phase Data
|
12 |
+
|
13 |
+
[](https://opensource.org/licenses/MIT)
|
14 |
+
[](https://arxiv.org/abs/2509.06888)
|
15 |
+
[](https://huggingface.co/collections/jhu-clsp/mmbert-a-modern-multilingual-encoder-68b725831d7c6e3acc435ed4)
|
16 |
+
[](https://github.com/jhu-clsp/mmBERT)
|
17 |
+
|
18 |
+
> **Phase 3 of 3**: Annealed language learning decay phase (100B tokens) with massive multilingual expansion to 1833 languages.
|
19 |
+
|
20 |
+
## π Data Composition
|
21 |
+
NOTE: there are multiple decay data mixtures: this mixture described below is the Decay-Cont mixture. However, the data in this repository is the Decay-Eng. If you are interested in the others, please let me know so I can prioritize it.
|
22 |
+
|
23 |
+
| Data Source | Tokens (B) | Percentage | Description |
|
24 |
+
|:------------|:-----------|:-----------|:------------|
|
25 |
+
| FineWeb2 | 78.5 | 76.0% | High-quality multilingual web crawl data |
|
26 |
+
| Wikipedia (MegaWika) | 9.5 | 9.2% | Encyclopedia articles (1833 languages) |
|
27 |
+
| Arxiv | 3.3 | 3.2% | Academic preprints |
|
28 |
+
| Textbooks (ProLong) | 3.1 | 3.0% | Educational content |
|
29 |
+
| Code (ProLong) | 2.8 | 2.7% | Code repositories and files |
|
30 |
+
| Books | 2.2 | 2.1% | Literature and reference books |
|
31 |
+
| DCLM (Dolmino) | 2.0 | 2.0% | High-quality English web data |
|
32 |
+
| Tulu Flan | 1.0 | 1.0% | Instruction-following data |
|
33 |
+
| Starcoder | 0.5 | 0.5% | Code repositories |
|
34 |
+
| Dolmino Math | 0.5 | 0.5% | Mathematical content |
|
35 |
+
| **Total** | **103.3** | **100.0%** | Optimized for rapid language acquisition |
|
36 |
+
|
37 |
+
## π Massive Language Coverage
|
38 |
+
|
39 |
+
This phase dramatically expands language coverage to **1833 languages**, implementing the novel **Cascading Annealed Language Learning (ALL)** approach:
|
40 |
+
|
41 |
+
- **Temperature Schedule**: Ο=0.3 (most uniform sampling)
|
42 |
+
- **Low-resource Focus**: Includes 1723 new languages with minimal data
|
43 |
+
- **Rapid Learning**: Demonstrates 68% performance improvement on Tigray and 26% on Faroese
|
44 |
+
- **Script Diversity**: Covers virtually all writing systems in FineWeb2
|
45 |
+
|
46 |
+
### Key Innovation: Annealed Language Learning
|
47 |
+
|
48 |
+
Rather than training on all languages simultaneously, MMBERT uses a cascading approach:
|
49 |
+
1. **Phase 1**: 60 high-resource languages (Ο=0.7)
|
50 |
+
2. **Phase 2**: 110 languages including mid-resource (Ο=0.5)
|
51 |
+
3. **Phase 3**: 1833 languages with focus on low-resource (Ο=0.3)
|
52 |
+
|
53 |
+
This enables rapid learning of new languages while maintaining performance on high-resource ones.
|
54 |
+
|
55 |
+
## βοΈ Key Features
|
56 |
+
|
57 |
+
- **Ultra-low Masking**: 5% mask rate for optimal learning efficiency
|
58 |
+
- **Model Merging**: Three decay variants (English-focused, 110-lang, 1833-lang) merged using TIES. This is the English focused version.
|
59 |
+
- **Quality Focus**: Emphasizes highest-quality data sources
|
60 |
+
|
61 |
+
## π Usage
|
62 |
+
|
63 |
+
For decay phase training, see the ModernBERT repo: https://github.com/AnswerDotAI/ModernBERT
|
64 |
+
|
65 |
+
### Direct Access
|
66 |
+
|
67 |
+
```python
|
68 |
+
from streaming import StreamingDataset
|
69 |
+
|
70 |
+
# Load the streaming dataset
|
71 |
+
dataset = StreamingDataset(
|
72 |
+
remote='https://huggingface.co/datasets/jhu-clsp/mmbert-decay',
|
73 |
+
local='/tmp/mmbert-decay-data',
|
74 |
+
shuffle=True
|
75 |
+
)
|
76 |
+
|
77 |
+
# Access samples
|
78 |
+
for sample in dataset:
|
79 |
+
text = sample['text']
|
80 |
+
# Process your data...
|
81 |
+
```
|
82 |
+
|
83 |
+
## π― Performance Impact
|
84 |
+
|
85 |
+
The decay phase demonstrates remarkable efficiency in low-resource language learning:
|
86 |
+
- **Tigray (TiQuAD)**: 68% improvement (12.1 F1 points) from including the language
|
87 |
+
- **Faroese (FoQA)**: 26% improvement (15.4 F1 points)
|
88 |
+
- **SOTA Performance**: Can even outperforms GPT-4o, Gemini 2.5 Pro
|
89 |
+
- **Rapid Acquisition**: Significant gains with only 100B tokens of exposure
|
90 |
+
|
91 |
+
## π Related Resources
|
92 |
+
|
93 |
+
- **Models**: [mmBERT Model Suite](https://huggingface.co/collections/jhu-clsp/mmbert-a-modern-multilingual-encoder-68b725831d7c6e3acc435ed4)
|
94 |
+
- **Phase 1**: [Pre-training Data](https://huggingface.co/datasets/jhu-clsp/mmbert-pretrain-p1-fineweb2-langs) (2.3T tokens)
|
95 |
+
- **Phase 2**: [Mid-training Data](https://huggingface.co/datasets/jhu-clsp/mmbert-midtraining) (600B tokens)
|
96 |
+
- **Checkpoints**: [Training Checkpoints](https://huggingface.co/datasets/jhu-clsp/mmbert-checkpoints)
|
97 |
+
- **Paper**: [Arxiv link](https://arxiv.org/abs/2509.06888)
|
98 |
+
- **Code**: [GitHub Repository](https://github.com/jhu-clsp/mmBERT)
|
99 |
+
|
100 |
+
## Citation
|
101 |
+
|
102 |
+
```bibtex
|
103 |
+
@misc{marone2025mmbertmodernmultilingualencoder,
|
104 |
+
title={mmBERT: A Modern Multilingual Encoder with Annealed Language Learning},
|
105 |
+
author={Marc Marone and Orion Weller and William Fleshman and Eugene Yang and Dawn Lawrie and Benjamin Van Durme},
|
106 |
+
year={2025},
|
107 |
+
eprint={2509.06888},
|
108 |
+
archivePrefix={arXiv},
|
109 |
+
primaryClass={cs.CL},
|
110 |
+
url={https://arxiv.org/abs/2509.06888},
|
111 |
+
}
|
112 |
+
```
|