GATE-AraBert-V1

This is GATE | General Arabic Text Embedding trained using SentenceTransformers in a multi-task setup. The system trains on the AllNLI and on the STS dataset.

Model Details

Model Description

• Model Type: Sentence Transformer
• Base model: Omartificial-Intelligence-Space/Arabic-Triplet-Matryoshka-V2
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 768 tokens
• Similarity Function: Cosine Similarity
• Training Datasets:
  • all-nli
  • sts
• Language: ar

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("Omartificial-Intelligence-Space/GATE-AraBert-v1")
# Run inference
sentences = [
    'الكلب البني مستلقي على جانبه على سجادة بيج، مع جسم أخضر في المقدمة.',
    'لقد مات الكلب',
    'شخص طويل القامة',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

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

Evaluation

Metrics

Semantic Similarity

• Dataset: sts-dev
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.8391
spearman_cosine	0.841
pearson_manhattan	0.8277
spearman_manhattan	0.8361
pearson_euclidean	0.8274
spearman_euclidean	0.8358
pearson_dot	0.8154
spearman_dot	0.818
pearson_max	0.8391
spearman_max	0.841

Semantic Similarity

• Dataset: sts-test
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.813
spearman_cosine	0.8173
pearson_manhattan	0.8114
spearman_manhattan	0.8164
pearson_euclidean	0.8103
spearman_euclidean	0.8158
pearson_dot	0.7908
spearman_dot	0.7887
pearson_max	0.813
spearman_max	0.8173

Acknowledgments

The author would like to thank Prince Sultan University for their invaluable support in this project. Their contributions and resources have been instrumental in the development and fine-tuning of these models.

## Citation

If you use the GATE, please cite it as follows:

@misc{nacar2025GATE,
      title={GATE: General Arabic Text Embedding for Enhanced Semantic Textual Similarity with Hybrid Loss Training}, 
      author={Omer Nacar, Anis Koubaa, Serry Taiseer Sibaee and Lahouari Ghouti},
      year={2025},
      note={Submitted to COLING 2025},
      url={https://huggingface.co./Omartificial-Intelligence-Space/GATE-AraBert-v1},
}