NeoAraBERT
We pretrain NeoAraBERT on diverse open-source and internal datasets covering modern standard, classical, and dialectal Arabic. We guided our design choices with Arabic tailored ablation studies including text normalization, light stemming, and diacritics-aware tokenization handling. We also performed POS-aware token masking and learning-rate scheduling ablation studies. We benchmarked NeoAraBERT against five top-performing Arabic models on 23 tasks, including a synonym-based task, Muradif, that directly assesses embedding quality with no additional fine-tuning. NeoAraBERT variants rank first in 18 tasks and improve average performance across the full benchmark suite.
This is the NeoAraBERT_Mix checkpoint, our best-performing checkpoint overall. This model was introduced at the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026). For more information, visit our website: https://acr.ps/neoarabert.
The available NeoAraBERT checkpoints:
| Model | Description | Link |
|---|---|---|
| NeoAraBERT (NeoAraBERT_Mix) | Trained on both Modern Standard Arabic and Dialectal Arabic. | this repository ✅ |
| NeoAraBERT_MSA | Trained on Modern Standard Arabic. | link |
| NeoAraBERT_DA | Trained on Dialectal Arabic. | link |
For detailed benchmarking, see https://acr.ps/neoarabert.
How to Use
Install these libraries:
pip install fast-disambig hf_transfer==0.1.9 torch==2.5.1 transformers==4.49.0 xformers==0.0.28.post3 torchvision torchaudio
Load the model and use it to generate embeddings:
from transformers import AutoModel, AutoTokenizer
model_name = "U4RASD/NeoAraBERT"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModel.from_pretrained(model_name, trust_remote_code=True)
# Tokenize input text
text = "المركز العربيّ للأبحاث ودراسة السياسات."
inputs = tokenizer(text, return_tensors="pt")
# Generate embeddings
outputs = model(**inputs)
embedding = outputs.last_hidden_state[:, 0, :]
print(embedding.shape)
Citation
If you use the code, model, or the Muradif benchmark, please cite:
@inproceedings{abou-chakra-etal-2026-neoarabert,
title = "{N}eo{A}ra{BERT}: A Modern Foundation Model for {A}rabic Embeddings with Diacritics-Aware Tokenization and {POS}-Targeted Masking",
author = "Abou Chakra, Chadi and
Hamoud, Hadi Khaled and
Rakan Al Mraikhat, Osama and
Abu Obaida, Qusai and
Ballout, Mohamad and
Zaraket, Fadi",
editor = "Liakata, Maria and
Moreira, Viviane P. and
Zhang, Jiajun and
Jurgens, David",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-acl.1293/",
doi = "10.18653/v1/2026.findings-acl.1293",
pages = "25952--25968",
ISBN = "979-8-89176-395-1",
abstract = "We present NeoAraBERT, a state-of-the-art open-source Arabic text-embedding model built on the NeoBERT architecture. We pre-train NeoAraBERT on diverse open-source and internal datasets covering modern standard, classical, and dialectal Arabic. We guided our design choices with Arabic tailored ablation studies including text normalization, light stemming, and diacritics-aware tokenization handling. We also performed more general POS-aware token masking and learning-rate scheduling ablation studies. We benchmarked NeoAraBERT against five top-performing Arabic models on 23 tasks, including a novel synonym-based task, ``Muradif'', that directly assesses embedding quality with no additional fine-tuning. NeoAraBERT variants (MSA, dialectal, and mixed) rank first in 18 tasks, second in two, third in two, and fourth in one task. They show strong performance on classical and modern standard Arabic, substantial margins of improvement ($>$7{\%}) in two tasks, and a $+$2.75{\%} improvement on average across all tasks. Our code and links to checkpoints for our model variants are available on our website: \url{https://acr.ps/neoarabert}."
}
Acknowledgements
We would like to acknowledge Ahmad Talal Salman from Assafir and Professor Amer Abdo Mouawad from the American University of Beirut for sharing Assafir data, which was instrumental to the work presented in this paper.
License
This model is licensed under the CC BY-SA 4.0 license. The text of the license can be found here.