README.md

---
tags:
- sentence-transformers
- sentence-similarity
- feature-extraction
- generated_from_trainer
- dataset_size:2280319
- loss:MatryoshkaLoss
- loss:MultipleNegativesRankingLoss
- Arabic
- EuroBert
- Semantic
base_model: EuroBERT/EuroBERT-610m
widget:
- source_sentence: امرأة شقراء تطل على مشهد (سياتل سبيس نيدل)
  sentences:
  - رجل يستمتع بمناظر جسر البوابة الذهبية
  - فتاة بالخارج تلعب في الثلج
  - شخص ما يأخذ في نظرة إبرة الفضاء.
- source_sentence: سوق الشرق الأوسط
  sentences:
  - مسرح أمريكي
  - متجر في الشرق الأوسط
  - البالغون صغار
- source_sentence: رجلين يتنافسان في ملابس فنون الدفاع عن النفس
  sentences:
  - هناك العديد من الناس الحاضرين.
  - الكلب الأبيض على الشاطئ
  - هناك شخص واحد فقط موجود.
- source_sentence: مجموعة من الناس تمشي بجانب شاحنة.
  sentences:
  - الناس يقفون
  - بعض الناس بالخارج
  - بعض الرجال يقودون على الطريق
- source_sentence: لاعبة كرة ناعمة ترمي الكرة إلى زميلتها في الفريق
  sentences:
  - شخصان يلعبان كرة البيسبول
  - الرجل ينظف
  - لاعبين لكرة البيسبول يجلسان على مقعد
pipeline_tag: feature-extraction
library_name: sentence-transformers
metrics:
- pearson_cosine
- spearman_cosine
model-index:
- name: SentenceTransformer based on EuroBERT/EuroBERT-610m
  results:
  - task:
      type: semantic-similarity
      name: Semantic Similarity
    dataset:
      name: sts dev 1152
      type: sts-dev-1152
    metrics:
    - type: pearson_cosine
      value: 0.8264443610084379
      name: Pearson Cosine
    - type: spearman_cosine
      value: 0.8306689922163598
      name: Spearman Cosine
  - task:
      type: semantic-similarity
      name: Semantic Similarity
    dataset:
      name: sts dev 960
      type: sts-dev-960
    metrics:
    - type: pearson_cosine
      value: 0.8258566703064338
      name: Pearson Cosine
    - type: spearman_cosine
      value: 0.830247434103489
      name: Spearman Cosine
  - task:
      type: semantic-similarity
      name: Semantic Similarity
    dataset:
      name: sts dev 768
      type: sts-dev-768
    metrics:
    - type: pearson_cosine
      value: 0.8243518007889306
      name: Pearson Cosine
    - type: spearman_cosine
      value: 0.8293091429698137
      name: Spearman Cosine
  - task:
      type: semantic-similarity
      name: Semantic Similarity
    dataset:
      name: sts dev 512
      type: sts-dev-512
    metrics:
    - type: pearson_cosine
      value: 0.8238103920299558
      name: Pearson Cosine
    - type: spearman_cosine
      value: 0.8293245725151981
      name: Spearman Cosine
license: mit
language:
- ar
---

# Ara-EuroBERT: Arabic Semantic Text Embeddings

<img src="https://i.ibb.co/d4svDscP/Clear-Familiar-situations-that-you-already-have-best-practices-for-4.png" width="150" align="left"/>

**Ara-EuroBERT-610M** is a [sentence-transformers](https://www.SBERT.net) model fine-tuned from [EuroBERT/EuroBERT-610m](https://huggingface.co/EuroBERT/EuroBERT-610m) specifically optimized for **Semantic Arabic text embeddings**.

This model maps sentences and paragraphs to a **1152-dimensional dense vector space**  and  **Maximum Sequence Length: 8,192 tokens**.


It can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

You can find more information on the base model at https://huggingface.co/EuroBERT/EuroBERT-610m

<br clear="left"/>

## Model Details & Benchmark Performance

<img src="https://cdn-uploads.huggingface.co/production/uploads/628f7a71dd993507cfcbe587/Kv78q7NmI3hhOXkRv30s9.png" width="1000" align="center"/>

The benchmark results above demonstrate the significant performance improvements of AraEuroBERT models compared to standard EuroBERT models:

- **STS17 Benchmark**: AraEuroBERT-610M achieves a score of 83, significantly outperforming the standard EuroBERT-610M (14) and even the much larger EuroBERT-2.1B (12).
- **STS22.v2 Benchmark**: AraEuroBERT-210M scores 61, outperforming both the larger AraEuroBERT-610M (53) and all standard EuroBERT variants.

These results highlight the effectiveness of our specialized fine-tuning for Arabic text embeddings, with even our smaller 210M parameter model demonstrating superior performance on Arabic semantic tasks.

### Metrics

#### Semantic Similarity

* Datasets: `sts-dev-1152`, `sts-dev-960`, `sts-dev-768` and `sts-dev-512`
* Evaluated with [<code>EmbeddingSimilarityEvaluator</code>](https://sbert.net/docs/package_reference/sentence_transformer/evaluation.html#sentence_transformers.evaluation.EmbeddingSimilarityEvaluator)

| Metric              | sts-dev-1152 | sts-dev-960 | sts-dev-768 | sts-dev-512 |
|:--------------------|:-------------|:------------|:------------|:------------|
| pearson_cosine      | 0.8264       | 0.8259      | 0.8244      | 0.8238      |
| **spearman_cosine** | **0.8307**   | **0.8302**  | **0.8293**  | **0.8293**  |


### Model Description
- **Model Type:** Sentence Transformer
- **Base model:** [EuroBERT/EuroBERT-610m](https://huggingface.co/EuroBERT/EuroBERT-610m)
- **Maximum Sequence Length:** 8192 tokens
- **Output Dimensionality:** 1152 dimensions (supports nested dimensions: 1152, 960, 768, 512)
- **Similarity Function:** Cosine Similarity
- **Training Dataset:** 2.28M training samples with Arabic text triplets
- **Language:** Arabic


### Full Model Architecture

```
SentenceTransformer(
  (0): Transformer({'max_seq_length': 8192, 'do_lower_case': False}) with Transformer model: EuroBertModel 
  (1): Pooling({'word_embedding_dimension': 1152, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
)
```

## Usage

### Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

```bash
pip install -U sentence-transformers
```

Then you can load this model and run inference:

```python
from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("Omartificial-Intelligence-Space/AraEuroBert-610M")

# Run inference
sentences = [
    'لاعبة كرة ناعمة ترمي الكرة إلى زميلتها في الفريق',
    'شخصان يلعبان كرة البيسبول',
    'لاعبين لكرة البيسبول يجلسان على مقعد',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 1152]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]
```

## Nested Embeddings

This model supports Matryoshka (nested) embeddings with the following dimensions:
- Full dimension: 1152
- Reduced dimensions: 960, 768, 512

You can select the embedding dimension based on your specific requirements, trading off between performance and computational efficiency.


## Citation

If you use this model in your research, please cite the following works:

```bibtex
@misc{boizard2025eurobertscalingmultilingualencoders,
      title={EuroBERT: Scaling Multilingual Encoders for European Languages}, 
      author={Nicolas Boizard and Hippolyte Gisserot-Boukhlef and Duarte M. Alves and André Martins and Ayoub Hammal and Caio Corro and Céline Hudelot and Emmanuel Malherbe and Etienne Malaboeuf and Fanny Jourdan and Gabriel Hautreux and João Alves and Kevin El-Haddad and Manuel Faysse and Maxime Peyrard and Nuno M. Guerreiro and Patrick Fernandes and Ricardo Rei and Pierre Colombo},
      year={2025},
      eprint={2503.05500},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.05500}, 
}
```

```bibtex
@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}
```

```bibtex
@misc{kusupati2024matryoshka,
    title={Matryoshka Representation Learning},
    author={Aditya Kusupati and Gantavya Bhatt and Aniket Rege and Matthew Wallingford and Aditya Sinha and Vivek Ramanujan and William Howard-Snyder and Kaifeng Chen and Sham Kakade and Prateek Jain and Ali Farhadi},
    year={2024},
    eprint={2205.13147},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}
```