ARA Reranker V1

🚀 Introducing ARM-V1 | Arabic Reranker Model (Version 1)

ARM-V1 is a specialized model for Arabic language reranking tasks, offering precise query - passage evaluation and interpretable relevance scores.

For more info please refer to this blog: ARM | Arabic Reranker Model.

✨ Features

• This model is designed specifically for Arabic language reranking tasks, optimized to handle queries and passages with precision.
• Unlike embedding models, which generate vector representations, this reranker directly evaluates the similarity between a question and a document, outputting a relevance score.
• Trained on a combination of positive and hard negative query - passage pairs, it excels in identifying the most relevant results.
• The output score can be transformed into a [0, 1] range using a sigmoid function, providing a clear and interpretable measure of relevance.

📚 Documentation

Arabic RAG Pipeline

📦 Installation

pip install sentence-transformers

💻 Usage Examples

Basic Usage

from sentence_transformers import CrossEncoder

# Load the cross-encoder model

# Define a query and a set of candidates with varying degrees of relevance
query = "تطبيقات الذكاء الاصطناعي تُستخدم في مختلف المجالات لتحسين الكفاءة."

# Candidates with varying relevance to the query
candidates = [
    "الذكاء الاصطناعي يساهم في تحسين الإنتاجية في الصناعات المختلفة.", # Highly relevant
    "نماذج التعلم الآلي يمكنها التعرف على الأنماط في مجموعات البيانات الكبيرة.", # Moderately relevant
    "الذكاء الاصطناعي يساعد الأطباء في تحليل الصور الطبية بشكل أفضل.", # Somewhat relevant
    "تستخدم الحيوانات التمويه كوسيلة للهروب من الحيوانات المفترسة.", # Irrelevant
]

# Create pairs of (query, candidate) for each candidate
query_candidate_pairs = [(query, candidate) for candidate in candidates]

# Get relevance scores from the model
scores = model.predict(query_candidate_pairs)

# Combine candidates with their scores and sort them by score in descending order (higher score = higher relevance)
ranked_candidates = sorted(zip(candidates, scores), key=lambda x: x[1], reverse=True)

# Output the ranked candidates with their scores
print("Ranked candidates based on relevance to the query:")
for i, (candidate, score) in enumerate(ranked_candidates, 1):
    print(f"Rank {i}:")
    print(f"Candidate: {candidate}")
    print(f"Score: {score}\n")

🔧 Technical Details

Evaluation

Dataset

• Size: 3000 samples.

Structure

Evaluation Process

• Query Grouping: Queries are grouped to evaluate the model's ability to rank candidate documents correctly for each query.
• Model Prediction: Each model predicts relevance scores for all candidate documents corresponding to a query.
• Metrics Calculation: Metrics are computed to measure how well the model ranks relevant documents higher than irrelevant ones.

📄 License

This project is licensed under the Apache-2.0 license.

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{nacar2025ARM,
      title={ARM, Arabic Reranker Model}, 
      author={Omer Nacar},
      year={2025},
      url={https://huggingface.co/Omartificial-Intelligence-Space/ARA-Reranker-V1},
}