Mizan Rerank V1

🚀 Mizan-Rerank-v1

A revolutionary open-source model for reranking Arabic long texts with exceptional efficiency and accuracy.

🚀 Quick Start

Mizan-Rerank-v1 is a leading open-source model based on the Transformer architecture, specifically designed for reranking search results in Arabic texts. With only 149 million parameters, it offers a perfect balance between performance and efficiency, outperforming larger models while using significantly fewer resources.

✨ Features

• Lightweight & Efficient: 149M parameters vs competitors with 278 - 568M parameters
• Long Text Processing: Handles up to 8192 tokens with sliding window technique
• High-Speed Inference: 3x faster than comparable models
• Arabic Language Optimization: Specifically fine-tuned for Arabic language nuances
• Resource Efficient: 75% less memory consumption than competitors

📊 Performance Benchmarks

Hardware Performance (RTX 4090 24GB)

MIRACL Dataset Results (ndcg@10)

Reranking and Triplet Datasets (ndcg@10)

🔧 Technical Details

Mizan-Rerank-v1 was trained on a diverse corpus of 741,159,981 tokens from:

• Authentic Arabic open-source datasets
• Manually crafted and processed text
• Purpose-generated synthetic data

This comprehensive training approach enables deep understanding of Arabic linguistic contexts.

📚 Documentation

How It Works

1. Query reception: The model receives a user query and candidate texts
2. Content analysis: Analyzes semantic relationships between query and each text
3. Relevance scoring: Assigns a relevance score to each text
4. Reranking: Sorts results by descending relevance score

💻 Usage Examples

Basic Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer

# Load model and tokenizer
model = AutoModelForSequenceClassification.from_pretrained("ALJIACHI/Mizan-Rerank-v1")
tokenizer = AutoTokenizer.from_pretrained("ALJIACHI/Mizan-Rerank-v1")

# Function to calculate relevance score
def get_relevance_score(query, passage):
    inputs = tokenizer(query, passage, return_tensors="pt", padding=True, truncation=True, max_length=8192)
    outputs = model(**inputs)
    return outputs.logits.item()

# Example usage
query = "ما هو تفسير الآية وجعلنا من الماء كل شيء حي"
passages = [
    "تعني الآية أن الماء هو عنصر أساسي في حياة جميع الكائنات الحية، وهو ضروري لاستمرار الحياة.",
    "تم اكتشاف كواكب خارج المجموعة الشمسية تحتوي على مياه متجمدة.",
    "تحدث القرآن الكريم عن البرق والرعد في عدة مواضع مختلفة."
]

# Get scores for each passage
scores = [(passage, get_relevance_score(query, passage)) for passage in passages]

# Rerank passages
reranked_passages = sorted(scores, key=lambda x: x[1], reverse=True)

# Print results
for passage, score in reranked_passages:
    print(f"Score: {score:.4f} | Passage: {passage}")

Practical Examples

Example 1

Question: What is the new tax law in 2024?

Example 2

Question: What is the interpretation of the verse "And We made from water every living thing"?

Example 3

Question: What are the benefits of vitamin D?

📈 Applications

Mizan-Rerank-v1 opens new horizons for Arabic NLP applications:

• Specialized Arabic search engines
• Archiving systems and digital libraries
• Conversational AI applications
• E-learning platforms
• Information retrieval systems

📝 Citation

If you use Mizan-Rerank-v1 in your research, please cite:

@software{Mizan_Rerank_v1_2025,
  author = {Ali Aljiachi},
  title = {Mizan-Rerank-v1: A Revolutionary Arabic Text Reranking Model},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/Aljiachi/Mizan-Rerank-v1}
}

@misc{modernbert,
      title={Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference}, 
      author={Benjamin Warner and Antoine Chaffin and Benjamin Clavié and Orion Weller and Oskar Hallström and Said Taghadouini and Alexis Gallagher and Raja Biswas and Faisal Ladhak and Tom Aarsen and Nathan Cooper and Griffin Adams and Jeremy Howard and Iacopo Poli},
      year={2024},
      eprint={2412.13663},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.13663}, 
}

📄 License

We release the Mizan-Rerank model model weights under the Apache 2.0 license.