Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,482 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
datasets:
|
4 |
+
- jhu-clsp/mmbert-decay
|
5 |
+
- jhu-clsp/mmbert-midtraining
|
6 |
+
- jhu-clsp/mmbert-pretrain-p1-fineweb2-langs
|
7 |
+
- jhu-clsp/mmbert-pretrain-p2-fineweb2-remaining
|
8 |
+
- jhu-clsp/mmbert-pretrain-p3-others
|
9 |
+
pipeline_tag: fill-mask
|
10 |
+
---
|
11 |
+
|
12 |
+
# mmBERT: A Modern Multilingual Encoder
|
13 |
+
|
14 |
+
[](https://opensource.org/licenses/MIT)
|
15 |
+
[](https://arxiv.org/abs/2509.06888)
|
16 |
+
[](https://huggingface.co/jhu-clsp/mmBERT-base)
|
17 |
+
[](https://huggingface.co/collections/jhu-clsp/mmbert-a-modern-multilingual-encoder-68b725831d7c6e3acc435ed4)
|
18 |
+
[](https://github.com/jhu-clsp/mmBERT)
|
19 |
+
|
20 |
+
> TL;DR: A state-of-the-art multilingual encoder trained on 3T+ tokens across 1800+ languages, introducing novel techniques for learning low-resource languages during the decay phase.
|
21 |
+
|
22 |
+
mmBERT is a modern multilingual encoder that significantly outperforms previous generation models like XLM-R on classification, embedding, and retrieval tasks. Built on the ModernBERT architecture with novel multilingual training innovations, mmBERT demonstrates that low-resource languages can be effectively learned during the decay phase of training. It is also significantly faster than any previous multilingual encoder.
|
23 |
+
|
24 |
+
## Table of Contents
|
25 |
+
- [Highlights](#highlights)
|
26 |
+
- [Quick Start](#quick-start)
|
27 |
+
- [Model Description](#model-description)
|
28 |
+
- [Novel Training Innovations](#novel-training-innovations)
|
29 |
+
- [Model Family](#model-family)
|
30 |
+
- [Training Data](#training-data)
|
31 |
+
- [Usage Examples](#usage-examples)
|
32 |
+
- [Fine-tuning Examples](#fine-tuning-examples)
|
33 |
+
- [Model Architecture](#model-architecture)
|
34 |
+
- [Citation](#citation)
|
35 |
+
|
36 |
+
|
37 |
+
## Quick Start
|
38 |
+
|
39 |
+
### Installation
|
40 |
+
```bash
|
41 |
+
pip install torch>=1.9.0
|
42 |
+
pip install transformers>=4.21.0
|
43 |
+
```
|
44 |
+
|
45 |
+
### Usage
|
46 |
+
|
47 |
+
```python
|
48 |
+
from transformers import AutoTokenizer, AutoModel
|
49 |
+
|
50 |
+
tokenizer = AutoTokenizer.from_pretrained("jhu-clsp/mmBERT-base")
|
51 |
+
model = AutoModel.from_pretrained("jhu-clsp/mmBERT-base")
|
52 |
+
|
53 |
+
inputs = tokenizer("Hello world", return_tensors="pt")
|
54 |
+
outputs = model(**inputs)
|
55 |
+
```
|
56 |
+
|
57 |
+
## Model Description
|
58 |
+
|
59 |
+
mmBERT represents the first significant advancement over XLM-R for massively multilingual encoder models. Key features include:
|
60 |
+
|
61 |
+
1. **Massive Language Coverage** - Trained on over 1800 languages with progressive inclusion strategy
|
62 |
+
2. **Modern Architecture** - Built on ModernBERT foundation with Flash Attention 2 and unpadding techniques
|
63 |
+
3. **Novel Training Recipe** - Introduces inverse mask scheduling and temperature sampling
|
64 |
+
4. **Open Training Data** - Complete 3T+ token dataset publicly available
|
65 |
+
5. **Decay Phase Innovation** - Demonstrates effective learning of low-resource languages in final training phase
|
66 |
+
|
67 |
+
The model uses bidirectional attention with masked language modeling objectives, optimized specifically for multilingual understanding and cross-lingual transfer.
|
68 |
+
|
69 |
+
## Novel Training Innovations
|
70 |
+
|
71 |
+
**Progressive Language Addition**: Start with 60 high-resource languages, expand to 110 mid-resource languages, then include all 1833 languages in decay phase.
|
72 |
+
|
73 |
+
**Inverse Mask Schedule**: Reduce mask ratio from 30% → 15% → 5% across training phases for progressively refined learning.
|
74 |
+
|
75 |
+
**Inverse Temperature Sampling**: Adjust multilingual sampling from high-resource bias (τ=0.7) to uniform sampling (τ=0.3).
|
76 |
+
|
77 |
+
**Model Merging**: Combine English-focused, high-resource, and all-language decay variants using TIES merging.
|
78 |
+
|
79 |
+
## Model Family
|
80 |
+
|
81 |
+
| Model | Total Params | Non-embed Params | Languages | Download |
|
82 |
+
|:------|:-------------|:------------------|:----------|:---------|
|
83 |
+
| [mmBERT-small](https://huggingface.co/jhu-clsp/mmBERT-small) | 140M | 42M | 1800+ | [](https://huggingface.co/jhu-clsp/mmBERT-small) |
|
84 |
+
| [mmBERT-base](https://huggingface.co/jhu-clsp/mmBERT-base) | 307M | 110M | 1800+ | [](https://huggingface.co/jhu-clsp/mmBERT-base) |
|
85 |
+
|
86 |
+
## Training Data
|
87 |
+
|
88 |
+
mmBERT training data is publicly available across different phases:
|
89 |
+
|
90 |
+
| Phase | Dataset | Tokens | Description |
|
91 |
+
|:------|:--------|:-------|:------------|
|
92 |
+
| Pre-training P1 | [mmbert-pretrain-p1](https://huggingface.co/datasets/jhu-clsp/mmbert-pretrain-p1-fineweb2-langs) | 2.3T | 60 languages, foundational training |
|
93 |
+
| Pre-training P2 | [mmbert-pretrain-p2](https://huggingface.co/datasets/jhu-clsp/mmbert-pretrain-p2-fineweb2-langs) | - | Extension data for pre-training phase |
|
94 |
+
| Pre-training P3 | [mmbert-pretrain-p3](https://huggingface.co/datasets/jhu-clsp/mmbert-pretrain-p3-fineweb2-langs) | - | Final pre-training data |
|
95 |
+
| Mid-training | [mmbert-midtraining](https://huggingface.co/datasets/jhu-clsp/mmbert-midtraining-data) | 600B | 110 languages, context extension to 8K |
|
96 |
+
| Decay Phase | [mmbert-decay](https://huggingface.co/datasets/jhu-clsp/mmbert-decay-data) | 100B | 1833 languages, premium quality |
|
97 |
+
|
98 |
+
**Data Sources**: Filtered DCLM (English), FineWeb2 (multilingual), FineWeb2-HQ (20 high-resource languages), Wikipedia (MegaWika), code repositories (StarCoder, ProLong), academic papers (ArXiv, PeS2o), and community discussions (StackExchange).
|
99 |
+
|
100 |
+
## Model Architecture
|
101 |
+
|
102 |
+
| Parameter | mmBERT-small | mmBERT-base |
|
103 |
+
|:----------|:-------------|:------------|
|
104 |
+
| Layers | 22 | 22 |
|
105 |
+
| Hidden Size | 384 | 768 |
|
106 |
+
| Intermediate Size | 1152 | 1152 |
|
107 |
+
| Attention Heads | 6 | 12 |
|
108 |
+
| Total Parameters | 140M | 307M |
|
109 |
+
| Non-embedding Parameters | 42M | 110M |
|
110 |
+
| Max Sequence Length | 8192 | 8192 |
|
111 |
+
| Vocabulary Size | 256,000 | 256,000 |
|
112 |
+
| Tokenizer | Gemma 2 | Gemma 2 |
|
113 |
+
|
114 |
+
## Usage Examples
|
115 |
+
|
116 |
+
### Masked Language Modeling
|
117 |
+
|
118 |
+
```python
|
119 |
+
from transformers import AutoTokenizer, AutoModelForMaskedLM
|
120 |
+
import torch
|
121 |
+
|
122 |
+
tokenizer = AutoTokenizer.from_pretrained("jhu-clsp/mmBERT-base")
|
123 |
+
model = AutoModelForMaskedLM.from_pretrained("jhu-clsp/mmBERT-base")
|
124 |
+
|
125 |
+
def predict_masked_token(text):
|
126 |
+
inputs = tokenizer(text, return_tensors="pt")
|
127 |
+
with torch.no_grad():
|
128 |
+
outputs = model(**inputs)
|
129 |
+
|
130 |
+
mask_indices = torch.where(inputs["input_ids"] == tokenizer.mask_token_id)
|
131 |
+
predictions = outputs.logits[mask_indices]
|
132 |
+
top_tokens = torch.topk(predictions, 5, dim=-1)
|
133 |
+
|
134 |
+
return [tokenizer.decode(token) for token in top_tokens.indices[0]]
|
135 |
+
|
136 |
+
# Works across languages
|
137 |
+
texts = [
|
138 |
+
"The capital of France is [MASK].",
|
139 |
+
"La capital de España es [MASK].",
|
140 |
+
"Die Hauptstadt von Deutschland ist [MASK]."
|
141 |
+
]
|
142 |
+
|
143 |
+
for text in texts:
|
144 |
+
predictions = predict_masked_token(text)
|
145 |
+
print(f"Text: {text}")
|
146 |
+
print(f"Predictions: {predictions}")
|
147 |
+
```
|
148 |
+
|
149 |
+
### Cross-lingual Embeddings
|
150 |
+
|
151 |
+
```python
|
152 |
+
from transformers import AutoTokenizer, AutoModel
|
153 |
+
import torch
|
154 |
+
from sklearn.metrics.pairwise import cosine_similarity
|
155 |
+
|
156 |
+
tokenizer = AutoTokenizer.from_pretrained("jhu-clsp/mmBERT-base")
|
157 |
+
model = AutoModel.from_pretrained("jhu-clsp/mmBERT-base")
|
158 |
+
|
159 |
+
def get_embeddings(texts):
|
160 |
+
inputs = tokenizer(texts, padding=True, truncation=True, return_tensors="pt")
|
161 |
+
|
162 |
+
with torch.no_grad():
|
163 |
+
outputs = model(**inputs)
|
164 |
+
embeddings = outputs.last_hidden_state.mean(dim=1)
|
165 |
+
|
166 |
+
return embeddings.numpy()
|
167 |
+
|
168 |
+
multilingual_texts = [
|
169 |
+
"Artificial intelligence is transforming technology",
|
170 |
+
"La inteligencia artificial está transformando la tecnología",
|
171 |
+
"L'intelligence artificielle transforme la technologie",
|
172 |
+
"人工智能正在改变技术"
|
173 |
+
]
|
174 |
+
|
175 |
+
embeddings = get_embeddings(multilingual_texts)
|
176 |
+
similarities = cosine_similarity(embeddings)
|
177 |
+
print("Cross-lingual similarity matrix:")
|
178 |
+
print(similarities)
|
179 |
+
```
|
180 |
+
|
181 |
+
## Fine-tuning Examples
|
182 |
+
|
183 |
+
### Dense Retrieval with Sentence Transformers
|
184 |
+
|
185 |
+
<details>
|
186 |
+
<summary>Click to expand dense retrieval fine-tuning example</summary>
|
187 |
+
|
188 |
+
```python
|
189 |
+
import argparse
|
190 |
+
from datasets import load_dataset
|
191 |
+
from sentence_transformers import (
|
192 |
+
SentenceTransformer,
|
193 |
+
SentenceTransformerTrainer,
|
194 |
+
SentenceTransformerTrainingArguments,
|
195 |
+
)
|
196 |
+
from sentence_transformers.evaluation import TripletEvaluator
|
197 |
+
from sentence_transformers.losses import CachedMultipleNegativesRankingLoss
|
198 |
+
from sentence_transformers.training_args import BatchSamplers
|
199 |
+
|
200 |
+
def main():
|
201 |
+
parser = argparse.ArgumentParser()
|
202 |
+
parser.add_argument("--lr", type=float, default=8e-5)
|
203 |
+
parser.add_argument("--model_name", type=str, default="jhu-clsp/mmBERT-base")
|
204 |
+
args = parser.parse_args()
|
205 |
+
|
206 |
+
lr = args.lr
|
207 |
+
model_name = args.model_name
|
208 |
+
model_shortname = model_name.split("/")[-1]
|
209 |
+
|
210 |
+
model = SentenceTransformer(model_name)
|
211 |
+
|
212 |
+
dataset = load_dataset(
|
213 |
+
"sentence-transformers/msmarco-co-condenser-margin-mse-sym-mnrl-mean-v1",
|
214 |
+
"triplet-hard",
|
215 |
+
split="train",
|
216 |
+
)
|
217 |
+
dataset_dict = dataset.train_test_split(test_size=1_000, seed=12)
|
218 |
+
train_dataset = dataset_dict["train"].select(range(1_250_000))
|
219 |
+
eval_dataset = dataset_dict["test"]
|
220 |
+
|
221 |
+
loss = CachedMultipleNegativesRankingLoss(model, mini_batch_size=16)
|
222 |
+
run_name = f"{model_shortname}-DPR-{lr}"
|
223 |
+
|
224 |
+
training_args = SentenceTransformerTrainingArguments(
|
225 |
+
output_dir=f"output/{model_shortname}/{run_name}",
|
226 |
+
num_train_epochs=1,
|
227 |
+
per_device_train_batch_size=512,
|
228 |
+
per_device_eval_batch_size=512,
|
229 |
+
warmup_ratio=0.05,
|
230 |
+
fp16=False,
|
231 |
+
bf16=True,
|
232 |
+
batch_sampler=BatchSamplers.NO_DUPLICATES,
|
233 |
+
learning_rate=lr,
|
234 |
+
save_strategy="steps",
|
235 |
+
save_steps=500,
|
236 |
+
save_total_limit=2,
|
237 |
+
logging_steps=500,
|
238 |
+
run_name=run_name,
|
239 |
+
)
|
240 |
+
|
241 |
+
dev_evaluator = TripletEvaluator(
|
242 |
+
anchors=eval_dataset["query"],
|
243 |
+
positives=eval_dataset["positive"],
|
244 |
+
negatives=eval_dataset["negative"],
|
245 |
+
name="msmarco-co-condenser-dev",
|
246 |
+
)
|
247 |
+
dev_evaluator(model)
|
248 |
+
|
249 |
+
trainer = SentenceTransformerTrainer(
|
250 |
+
model=model,
|
251 |
+
args=training_args,
|
252 |
+
train_dataset=train_dataset,
|
253 |
+
eval_dataset=eval_dataset,
|
254 |
+
loss=loss,
|
255 |
+
evaluator=dev_evaluator,
|
256 |
+
)
|
257 |
+
trainer.train()
|
258 |
+
|
259 |
+
model.save_pretrained(f"output/{model_shortname}/{run_name}/final")
|
260 |
+
model.push_to_hub(run_name, private=False)
|
261 |
+
|
262 |
+
if __name__ == "__main__":
|
263 |
+
main()
|
264 |
+
```
|
265 |
+
|
266 |
+
</details>
|
267 |
+
|
268 |
+
### Cross-lingual Classification
|
269 |
+
|
270 |
+
<details>
|
271 |
+
<summary>Click to expand multilingual classification fine-tuning example</summary>
|
272 |
+
|
273 |
+
```python
|
274 |
+
from transformers import (
|
275 |
+
AutoTokenizer,
|
276 |
+
AutoModelForSequenceClassification,
|
277 |
+
TrainingArguments,
|
278 |
+
Trainer
|
279 |
+
)
|
280 |
+
from datasets import load_dataset
|
281 |
+
import numpy as np
|
282 |
+
from sklearn.metrics import accuracy_score, f1_score
|
283 |
+
|
284 |
+
def compute_metrics(eval_pred):
|
285 |
+
predictions, labels = eval_pred
|
286 |
+
predictions = np.argmax(predictions, axis=1)
|
287 |
+
return {
|
288 |
+
'accuracy': accuracy_score(labels, predictions),
|
289 |
+
'f1': f1_score(labels, predictions, average='weighted')
|
290 |
+
}
|
291 |
+
|
292 |
+
def main():
|
293 |
+
model_name = "jhu-clsp/mmBERT-base"
|
294 |
+
|
295 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
296 |
+
model = AutoModelForSequenceClassification.from_pretrained(
|
297 |
+
model_name,
|
298 |
+
num_labels=3
|
299 |
+
)
|
300 |
+
|
301 |
+
dataset = load_dataset("xnli", "all_languages")
|
302 |
+
|
303 |
+
def tokenize_function(examples):
|
304 |
+
texts = [f"{p} {tokenizer.sep_token} {h}"
|
305 |
+
for p, h in zip(examples["premise"], examples["hypothesis"])]
|
306 |
+
|
307 |
+
return tokenizer(
|
308 |
+
texts,
|
309 |
+
truncation=True,
|
310 |
+
padding=True,
|
311 |
+
max_length=512
|
312 |
+
)
|
313 |
+
|
314 |
+
train_dataset = dataset["train"].map(tokenize_function, batched=True)
|
315 |
+
eval_dataset = dataset["validation"].map(tokenize_function, batched=True)
|
316 |
+
|
317 |
+
training_args = TrainingArguments(
|
318 |
+
output_dir="./mmbert-xnli",
|
319 |
+
learning_rate=3e-5,
|
320 |
+
per_device_train_batch_size=32,
|
321 |
+
per_device_eval_batch_size=32,
|
322 |
+
num_train_epochs=3,
|
323 |
+
weight_decay=0.01,
|
324 |
+
evaluation_strategy="epoch",
|
325 |
+
save_strategy="epoch",
|
326 |
+
load_best_model_at_end=True,
|
327 |
+
metric_for_best_model="f1",
|
328 |
+
greater_is_better=True,
|
329 |
+
)
|
330 |
+
|
331 |
+
trainer = Trainer(
|
332 |
+
model=model,
|
333 |
+
args=training_args,
|
334 |
+
train_dataset=train_dataset,
|
335 |
+
eval_dataset=eval_dataset,
|
336 |
+
compute_metrics=compute_metrics,
|
337 |
+
)
|
338 |
+
|
339 |
+
trainer.train()
|
340 |
+
|
341 |
+
if __name__ == "__main__":
|
342 |
+
main()
|
343 |
+
```
|
344 |
+
|
345 |
+
</details>
|
346 |
+
|
347 |
+
### Multilingual Reranking
|
348 |
+
|
349 |
+
<details>
|
350 |
+
<summary>Click to expand multilingual reranking fine-tuning example</summary>
|
351 |
+
|
352 |
+
```python
|
353 |
+
import logging
|
354 |
+
from datasets import load_dataset
|
355 |
+
from sentence_transformers.cross_encoder import (
|
356 |
+
CrossEncoder,
|
357 |
+
CrossEncoderModelCardData,
|
358 |
+
CrossEncoderTrainer,
|
359 |
+
CrossEncoderTrainingArguments,
|
360 |
+
)
|
361 |
+
from sentence_transformers.cross_encoder.evaluation import CrossEncoderNanoBEIREvaluator
|
362 |
+
from sentence_transformers.cross_encoder.losses import BinaryCrossEntropyLoss
|
363 |
+
from sentence_transformers.util import mine_hard_negatives
|
364 |
+
from sentence_transformers import SentenceTransformer
|
365 |
+
import torch
|
366 |
+
|
367 |
+
def main():
|
368 |
+
model_name = "jhu-clsp/mmBERT-base"
|
369 |
+
train_batch_size = 32
|
370 |
+
num_epochs = 2
|
371 |
+
num_hard_negatives = 7
|
372 |
+
|
373 |
+
model = CrossEncoder(
|
374 |
+
model_name,
|
375 |
+
model_card_data=CrossEncoderModelCardData(
|
376 |
+
language="multilingual",
|
377 |
+
license="mit",
|
378 |
+
),
|
379 |
+
)
|
380 |
+
|
381 |
+
full_dataset = load_dataset("sentence-transformers/gooaq", split="train").select(range(50_000))
|
382 |
+
dataset_dict = full_dataset.train_test_split(test_size=1_000, seed=42)
|
383 |
+
train_dataset = dataset_dict["train"]
|
384 |
+
eval_dataset = dataset_dict["test"]
|
385 |
+
|
386 |
+
embedding_model = SentenceTransformer("sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2", device="cpu")
|
387 |
+
hard_train_dataset = mine_hard_negatives(
|
388 |
+
train_dataset,
|
389 |
+
embedding_model,
|
390 |
+
num_negatives=num_hard_negatives,
|
391 |
+
margin=0,
|
392 |
+
range_min=0,
|
393 |
+
range_max=100,
|
394 |
+
sampling_strategy="top",
|
395 |
+
batch_size=2048,
|
396 |
+
output_format="labeled-pair",
|
397 |
+
use_faiss=True,
|
398 |
+
)
|
399 |
+
|
400 |
+
loss = BinaryCrossEntropyLoss(model=model, pos_weight=torch.tensor(num_hard_negatives))
|
401 |
+
|
402 |
+
nano_beir_evaluator = CrossEncoderNanoBEIREvaluator(
|
403 |
+
dataset_names=["msmarco", "nfcorpus", "nq"],
|
404 |
+
batch_size=train_batch_size,
|
405 |
+
)
|
406 |
+
|
407 |
+
args = CrossEncoderTrainingArguments(
|
408 |
+
output_dir="./mmbert-reranker",
|
409 |
+
num_train_epochs=num_epochs,
|
410 |
+
per_device_train_batch_size=train_batch_size,
|
411 |
+
per_device_eval_batch_size=train_batch_size,
|
412 |
+
learning_rate=2e-5,
|
413 |
+
warmup_ratio=0.1,
|
414 |
+
fp16=False,
|
415 |
+
bf16=True,
|
416 |
+
dataloader_num_workers=4,
|
417 |
+
load_best_model_at_end=True,
|
418 |
+
metric_for_best_model="eval_msmarco_ndcg@10",
|
419 |
+
eval_strategy="steps",
|
420 |
+
eval_steps=1000,
|
421 |
+
save_strategy="steps",
|
422 |
+
save_steps=1000,
|
423 |
+
save_total_limit=2,
|
424 |
+
logging_steps=200,
|
425 |
+
seed=42,
|
426 |
+
)
|
427 |
+
|
428 |
+
trainer = CrossEncoderTrainer(
|
429 |
+
model=model,
|
430 |
+
args=args,
|
431 |
+
train_dataset=hard_train_dataset,
|
432 |
+
loss=loss,
|
433 |
+
evaluator=nano_beir_evaluator,
|
434 |
+
)
|
435 |
+
trainer.train()
|
436 |
+
|
437 |
+
model.save_pretrained("./mmbert-reranker/final")
|
438 |
+
|
439 |
+
if __name__ == "__main__":
|
440 |
+
main()
|
441 |
+
```
|
442 |
+
|
443 |
+
</details>
|
444 |
+
|
445 |
+
## Training Data
|
446 |
+
|
447 |
+
mmBERT was trained on a carefully curated 3T+ token multilingual dataset:
|
448 |
+
|
449 |
+
| Phase | Dataset | Description |
|
450 |
+
|:------|:--------|:------------|
|
451 |
+
| [Pre-training P1](https://huggingface.co/datasets/jhu-clsp/mmbert-pretrain-p1-fineweb2-langs) | 2.3T tokens | 60 languages, diverse data mixture |
|
452 |
+
| [Pre-training P2](https://huggingface.co/datasets/jhu-clsp/mmbert-pretrain-p2-fineweb2-langs) | - | Extension data for pre-training |
|
453 |
+
| [Pre-training P3](https://huggingface.co/datasets/jhu-clsp/mmbert-pretrain-p3-fineweb2-langs) | - | Final pre-training data |
|
454 |
+
| [Mid-training](https://huggingface.co/datasets/jhu-clsp/mmbert-midtraining-data) | 600B tokens | 110 languages, context extension |
|
455 |
+
| [Decay Phase](https://huggingface.co/datasets/jhu-clsp/mmbert-decay-data) | 100B tokens | 1833 languages, premium quality |
|
456 |
+
|
457 |
+
**Primary Sources:**
|
458 |
+
- **Filtered DCLM**: High-quality English content
|
459 |
+
- **FineWeb2**: Broad multilingual web coverage (1800+ languages)
|
460 |
+
- **FineWeb2-HQ**: Filtered subset of 20 high-resource languages
|
461 |
+
- **Code**: StarCoder and ProLong repositories
|
462 |
+
- **Academic**: ArXiv papers and PeS2o scientific content
|
463 |
+
- **Reference**: Wikipedia (MegaWika) and textbooks
|
464 |
+
- **Community**: StackExchange discussions
|
465 |
+
|
466 |
+
|
467 |
+
## Citation
|
468 |
+
|
469 |
+
If you use mmBERT in your research, please cite our work:
|
470 |
+
|
471 |
+
```bibtex
|
472 |
+
@misc{marone2025mmbertmodernmultilingualencoder,
|
473 |
+
title={mmBERT: A Modern Multilingual Encoder with Annealed Language Learning},
|
474 |
+
author={Marc Marone and Orion Weller and William Fleshman and Eugene Yang and Dawn Lawrie and Benjamin Van Durme},
|
475 |
+
year={2025},
|
476 |
+
eprint={2509.06888},
|
477 |
+
archivePrefix={arXiv},
|
478 |
+
primaryClass={cs.CL},
|
479 |
+
url={https://arxiv.org/abs/2509.06888},
|
480 |
+
}
|
481 |
+
```
|
482 |
+
"""
|