README.md

metadata

license: mit
datasets:
  - yiyic/oscar_arb_Arab_train
  - yiyic/oscar_arb_Arab_test
  - SaiedAlshahrani/Arabic_Wikipedia_20230101_bots
  - ClusterlabAi/101_billion_arabic_words_dataset
language:
  - ar
metrics:
  - f1
  - exact_match
base_model:
  - answerdotai/ModernBERT-base
tags:
  - Embedding
  - Arabic
  - Sentiment_Analysis
  - QA
  - NER

Model Card: ModernAraBERT

Summary

• Arabic encoder adapted from answerdotai/ModernBERT-base via continued pretraining on Arabic corpora (~9.8GB).
• Strong results across SA, NER (Macro-F1), and QA EM vs. AraBERT/mBERT/MARBERT.
• License: MIT · Paper: LREC 2026 · Hub: gizadatateam/ModernAraBERT

Intended Uses

• Masked LM, feature extraction, and transfer learning for Arabic tasks.
• Downstream: sentiment analysis, NER, extractive QA, general classification/labeling.

How to use

from transformers import AutoTokenizer, AutoModelForMaskedLM
name = "gizadatateam/ModernAraBERT"
model = AutoModelForMaskedLM.from_pretrained(name)
tokenizer = AutoTokenizer.from_pretrained(name)

Training data and recipe (brief)

• Corpora: OSIAN, Arabic Billion Words, Arabic Wikipedia, OSCAR Arabic
• Tokenizer: ModernBERT vocab + 80K Arabic tokens
• Objective: MLM (3 epochs; 128→512 seq len)
• Hardware: A100 40GB; framework: PyTorch + Transformers + Accelerate

Evaluation (from paper)

Sentiment Analysis — Macro-F1 (%)

Model	LABR	HARD	AJGT
AraBERTv1	45.35	72.65	58.01
AraBERTv2	45.79	67.10	53.59
mBERT	44.18	71.70	61.55
MARBERT	45.54	67.39	60.63
ModernAraBERT	56.45	89.37	70.54

NER — Macro-F1 (%)

Model	Macro-F1
AraBERTv1	13.46
AraBERTv2	16.77
mBERT	12.15
MARBERT	7.42
ModernAraBERT	28.23

QA (ARCD test) — EM (%)

Model	EM
AraBERT	25.36
AraBERTv2	26.08
mBERT	25.12
MARBERT	23.58
ModernAraBERT	27.10

Citation

@inproceedings{<paper_id>,
  title={Efficient Adaptation of English Language Models for Low-Resource and Morphologically Rich Languages: The Case of Arabic},
  author={Maher, Eldamaty, Ashraf, ElShawi, Mostafa},
  booktitle={Proceedings of <conference_name>},
  year={2025},
  organization={<conference_name>}
}