DIMI Embedding model

State-of-the-art Arabic Sentence Embeddings

Model Description

This model is a Matryoshka representation learning version of AraBERT specifically fine-tuned for Arabic Natural Language Inference (NLI) tasks. It generates embeddings that can be truncated to different dimensions (768, 512, 256, 128, 64) while maintaining strong performance across all sizes.

The model is based on aubmindlab/bert-base-arabertv02 and trained using the Matryoshka Representation Learning approach, which allows for flexible embedding dimensions without retraining.

Key Features

  • 🔄 Flexible Dimensions: Single model supports embeddings of size 768, 512, 256, 128, and 64
  • 🚀 High Performance: Consistently outperforms base model across all dimensions
  • 📊 Arabic NLI Optimized: Specifically trained on Arabic Natural Language Inference tasks
  • Efficient: Smaller dimensions offer faster inference with minimal performance loss
  • 🎯 Binary Classification: Optimized for entailment vs contradiction classification

Performance Results

Our model shows significant improvements over the base AraBERT model across all embedding dimensions:

Dimension Matryoshka Accuracy Base Accuracy Matryoshka F1 Base F1 Improvement
768 80.3% 56.8% 81.15% 41.94% +39.21%
512 80.6% 56.9% 81.36% 44.32% +37.05%
256 80.95% 55.65% 81.42% 38.7% +42.72%
128 81.25% 56.7% 81.37% 40.6% +40.77%
64 81.0% 55.8% 80.51% 37.92% +42.59%

Quick Start

Installation

pip install sentence-transformers torch

Basic Usage

from sentence_transformers import SentenceTransformer

# Load the model
model = SentenceTransformer('AhmedZaky1/arabic-bert-nli-matryoshka')

# Example sentences
sentences = [
    "الطقس جميل اليوم",
    "إنه يوم مشمس وجميل", 
    "أحب قراءة الكتب"
]

# Generate embeddings (default: full 768 dimensions)
embeddings = model.encode(sentences)
print(f"Full embeddings shape: {embeddings.shape}")

# Use different dimensions by truncating
embeddings_256 = embeddings[:, :256]  # Use first 256 dimensions
embeddings_128 = embeddings[:, :128]  # Use first 128 dimensions
embeddings_64 = embeddings[:, :64]    # Use first 64 dimensions

print(f"256-dim embeddings shape: {embeddings_256.shape}")

Similarity Computation

from sentence_transformers import util

# Compute similarity between sentences
sentence1 = "القطة تجلس على السجادة"
sentence2 = "الكلب يلعب في الحديقة"

embeddings = model.encode([sentence1, sentence2])
similarity = util.cos_sim(embeddings[0], embeddings[1])
print(f"Similarity: {similarity.item():.4f}")

NLI Classification

def classify_nli_pair(premise, hypothesis, threshold=0.6):
    """
    Classify Natural Language Inference relationship

    Args:
        premise: The premise sentence
        hypothesis: The hypothesis sentence  
        threshold: Similarity threshold for classification

    Returns:
        str: 'entailment' if similarity > threshold, else 'contradiction'
    """
    embeddings = model.encode([premise, hypothesis])
    similarity = util.cos_sim(embeddings[0], embeddings[1]).item()

    return 'entailment' if similarity > threshold else 'contradiction'

# Example usage
premise = "الرجل يقرأ كتاباً في المكتبة"
hypothesis = "شخص يقرأ في مكان هادئ"

result = classify_nli_pair(premise, hypothesis)
print(f"Relationship: {result}")

Choosing the Right Dimension

  • 768 dimensions: Maximum accuracy for critical applications
  • 512 dimensions: Best balance of performance and efficiency
  • 256 dimensions: Good performance with 3x faster inference
  • 128 dimensions: Suitable for real-time applications
  • 64 dimensions: Ultra-fast inference for large-scale processing

Training Details

Dataset

  • Training Data: Arabic-NLI-Pair-Class dataset from Omartificial-Intelligence-Space
  • Language: Modern Standard Arabic (MSA)
  • Task Type: Binary classification (entailment vs contradiction)

Training Configuration

  • Base Model: aubmindlab/bert-base-arabertv02
  • Max Sequence Length: 75 tokens
  • Batch Size: 64
  • Epochs: 5
  • Matryoshka Dimensions: [768, 512, 256, 128, 64]
  • Loss Function: MatryoshkaLoss with CosineSimilarityLoss
  • Optimization: AdamW with automatic mixed precision (AMP)

Use Cases

  1. Arabic Text Similarity: Measure semantic similarity between Arabic texts
  2. Natural Language Inference: Determine logical relationships between Arabic sentences
  3. Information Retrieval: Find relevant Arabic documents based on queries
  4. Semantic Search: Build Arabic search engines with semantic understanding
  5. Text Classification: Use embeddings as features for downstream Arabic NLP tasks

Citation

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

@model{arabic-bert-nli-matryoshka,
  title={Arabic BERT NLI Matryoshka Embeddings},
  author={Ahmed Mouad},
  year={2025},
  url={https://huggingface.co/AhmedZaky1/arabic-bert-nli-matryoshka}
}

Acknowledgments

  • AraBERT Team: For the excellent base model (aubmindlab/bert-base-arabertv02)
  • Sentence Transformers: For the robust training framework
  • Matryoshka Representation Learning: For the innovative approach to nested embeddings
  • Arabic NLI Dataset: Omartificial-Intelligence-Space for the training data

License

This model is released under the Apache 2.0 License.


Model Version: 1.0
Last Updated: May 2025
Framework: sentence-transformers
Language: Arabic (العربية)

Downloads last month
92
Safetensors
Model size
135M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for AhmedZaky1/DIMI-embedding-matryoshka-arabic

Finetuned
(3993)
this model
Finetunes
1 model

Dataset used to train AhmedZaky1/DIMI-embedding-matryoshka-arabic

Evaluation results

  • Best Accuracy (128 dim) on Arabic NLI Pair Classification
    self-reported
    0.813
  • Best F1 (256 dim) on Arabic NLI Pair Classification
    self-reported
    0.814