AraModernBert-base-V1.0

Model Description

AraModernBERT is an advanced Arabic language model built on the ModernBERT architecture. This model represents a significant advancement in Arabic language understanding, combining state-of-the-art transformer design innovations with extensive training on 100 Gigabytes of Arabic text.

AraModernBERT was developed through the following process:

1. Custom Tokenizer Training: We built a specialized tokenizer optimized for Arabic language processing with a vocabulary size of 50,280 tokens.
2. Transtokenization: We used the transtokenization technique to optimally initialize the embedding layer for MLM (for more details, see the paper here).
3. Large-Scale Masked Language Modeling The model was trained on 100 Gigabytes of Arabic text.

Model Configuration

{
  "hidden_size": 768,
  "intermediate_size": 1152,
  "num_attention_heads": 12,
  "num_hidden_layers": 22,
  "max_position_embeddings": 8192,
  "vocab_size": 50280,
  "global_attn_every_n_layers": 3,
  "local_attention": 128,
  "global_rope_theta": 160000.0,
  "local_rope_theta": 10000.0,
  "architectures": ["ModernBertForMaskedLM"],
  "model_type": "modernbert",
  "cls_token_id": 3,
  "mask_token_id": 6,
  "pad_token_id": 5,
  "sep_token_id": 4,
  "unk_token_id": 2
}

Intended Uses & Limitations

AraModernBERT can be used for a wide range of Arabic NLP tasks, including:

• Text Embeddings & Representation
• Information Retrieval
• RAG (Retrieval Augmented Generation)
• Document Similarity
• Text Classification
• Sentiment Analysis

Limitations and Biases

• The model is optimized for Modern Standard Arabic and may show varying performance on dialectal Arabic variants or classical Arabic texts.
• Performance may vary across domains and specialized terminology.
• Users should be aware of potential biases present in the training data.

Evaluation Results

1. Semantic Textual Similarity (STS)

We fine-tuned the model on STS datasets to enhance semantic understanding capabilities:

• STS17: 0.831
• STS22: 0.617

Note: The STS-optimized model will be released soon as a separate checkpoint.

2. Text Classification

We finetuned AraModernBERT on a multi-class classification task using the SANAD dataset.

Overall Metrics:

• AraModernBERT:
  • Accuracy: 94.32%
  • F1 Score: 94.31%
  • Precision: 94.31%
  • Recall: 94.32%

Per-Class Performance (AraModernBERT):

3. Named Entity Recognition (NER)

The model achieved excellent performance on Arabic NER tasks:

• Accuracy: 90.39%
• Precision: 0.7357
• Recall: 0.7442
• F1: 0.7399

How to Use

Here's how to use AraModernBERT with the Transformers library:

from transformers import AutoTokenizer, AutoModel

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("NAMAA-Space/AraModernBERT-Base-V1.0")
model = AutoModel.from_pretrained("NAMAA-Space/AraModernBERT-Base-V1.0")

# Encode text
text = "مرحبا بكم في عالم الذكاء الاصطناعي"
inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)

# Get embeddings
embeddings = outputs.last_hidden_state

Masked Language Modeling Example

from transformers import AutoTokenizer, AutoModelForMaskedLM
import torch

tokenizer = AutoTokenizer.from_pretrained("NAMAA-Space/AraModernBERT-Base-V1.0")
model = AutoModelForMaskedLM.from_pretrained("NAMAA-Space/AraModernBERT-Base-V1.0")

text = "الذكاء الاصطناعي هو [MASK] المستقبل."
inputs = tokenizer(text, return_tensors="pt")
token_index = torch.where(inputs["input_ids"][0] == tokenizer.mask_token_id)[0]

outputs = model(**inputs)
predictions = outputs.logits
predicted_token_id = torch.argmax(predictions[0, token_index]).item()
predicted_token = tokenizer.decode(predicted_token_id)
print(predicted_token)  

Model Architecture

AraModernBERT inherits the modern architecture features from ModernBERT, while adding Trans-Tokenization approach:

• 22 transformer layers with 768 hidden dimensions
• Alternating Attention mechanism with global attention every 3 layers and a local attention window of 128 tokens
• Rotary Positional Embeddings (RoPE) with different theta values for global (160000.0) and local (10000.0) attention
• 8,192 token context window for processing longer documents
• Specialized vocabulary of 50,280 tokens optimized for Arabic

Technical Specifications

• Base Architecture: ModernBERT
• Parameters: ~149M (based on configuration)
• Context Length: 8,192 tokens
• Vocabulary Size: 50,280
• Hidden Size: 768
• Attention Heads: 12
• Hidden Layers: 22
• Intermediate Size: 1152

Citation

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

@inproceedings{elshehy-etal-2026-aramodernbert,
    title = "{A}ra{M}odern{BERT}: Transtokenized Initialization and Long-Context Encoder Modeling for {A}rabic",
    author = "Elshehy, Omar  and
      Nacar, Omer  and
      Djamai, Abdelbasset  and
      Ragab, Muhammed  and
      AL Jallad, Khloud  and
      Abdelazim, Mona",
    editor = "El-Haj, Mo  and
      Rayson, Paul  and
      Jarrar, Mustafa  and
      Ezeani, Ignatius  and
      Ezzini, Saad  and
      Ahmadi, Sina  and
      Haddad Haddad, Amal  and
      Amol, Cynthia  and
      Abdelali, Ahmad  and
      Abudalfa, Shadi",
    booktitle = "Proceedings of the 2nd Workshop on {NLP} for Languages Using {A}rabic Script",
    month = mar,
    year = "2026",
    address = "Rabat, Morocco",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.abjadnlp-1.39/",
    doi = "10.18653/v1/2026.abjadnlp-1.39",
    pages = "313--321",
}

Acknowledgements

This model builds upon the ModernBERT architecture developed by Answer.AI and LightOn. We acknowledge their contributions to the field of encoder-only models and extend their work to the Arabic language through our novel Trans-Tokenized approach.

@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}, 
}

@inproceedings{remy-delobelle2024transtokenization,
    title={Trans-Tokenization and Cross-lingual Vocabulary Transfers: Language Adaptation of {LLM}s for Low-Resource {NLP}},
    author={Remy, Fran{\c{c}}ois and Delobelle, Pieter and Avetisyan, Hayastan and Khabibullina, Alfiya and de Lhoneux, Miryam and Demeester, Thomas},
    booktitle={First Conference on Language Modeling},
    year={2024},
    url={https://openreview.net/forum?id=sBxvoDhvao}
}