Namaa-Reranker-v1

Namaa-Reranker-v1 🚀✨

NAMAA-space releases Namaa-Reranker-v1, a high-performance model fine-tuned on unicamp-dl/mmarco to elevate Arabic document retrieval and ranking to new heights! 📚🇸🇦

This model is designed to improve search relevance of arabic documents by accurately ranking documents based on their contextual fit for a given query.

Key Features 🔑

• Optimized for Arabic: Built on the highly performant Omartificial-Intelligence-Space/Arabic-Triplet-Matryoshka-V2 with exclusivly rich Arabic data.
• Advanced Document Ranking: Ranks results with precision, perfect for search engines, recommendation systems, and question-answering applications.
• State-of-the-Art Performance: Achieves excellent performance compared to famous rerankers(See Evaluation), ensuring reliable relevance and precision.

Example Use Cases 💼

• Retrieval Augmented Generation: Improve search result relevance for Arabic content.
• Content Recommendation: Deliver top-tier Arabic content suggestions.
• Question Answering: Boost answer retrieval quality in Arabic-focused systems.

Usage

Within sentence-transformers

The usage becomes easier when you have SentenceTransformers installed. Then, you can use the pre-trained models like this:

from sentence_transformers import CrossEncoder
model = CrossEncoder('NAMAA-Space/Namaa-Reranker-v1', max_length=512)

Query = 'كيف يمكن استخدام التعلم العميق في معالجة الصور الطبية؟'
Paragraph1 = 'التعلم العميق يساعد في تحليل الصور الطبية وتشخيص الأمراض'
Paragraph2 = 'الذكاء الاصطناعي يستخدم في تحسين الإنتاجية في الصناعات'

scores = model.predict([(Query, Paragraph1), (Query, Paragraph2)])

Evaluation

We evaluate our model on two different datasets using the metrics MAP, MRR and NDCG@10:

The purpose of this evaluation is to highlight the performance of our model with regards to: Relevant/Irrelevant labels and positive/multiple negatives documents:

Dataset 1: NAMAA-Space/Ar-Reranking-Eval

Dataset 2: NAMAA-Space/Arabic-Reranking-Triplet-5-Eval

As seen, The model performs extremly well in comparison to other famous rerankers.

WIP: More comparisons and evaluation on arabic datasets.