Research and Publications

At Kawn, we believe that building an advanced language ecosystem is incomplete without sharing knowledge and empowering the research community.

Pioneering Research Publications

Kuwain

We introduced an effective method for injecting Arabic into a small (1.5B) English-based model without compromising its prior knowledge. Kuwain showed an 8% improvement in Arabic performance and serves as the foundation for our specialized models.

link:🔗https://arxiv.org/abs/2504.15120

Sadeed

A lightweight and accurate model for diacritizing Arabic texts, built on Kuwain. It demonstrates strong performance compared to large, closed-source models and is accompanied by the SadeedDiac-25 benchmark.

link:🔗https://arxiv.org/abs/2504.21635

Mutarjim

An efficient model for bidirectional translation between Arabic and English, based on Kuwain. It surpasses larger models on several benchmarks and includes the Tarjama-25 benchmark for accurate evaluation.

link:🔗https://arxiv.org/abs/2505.17894

Lahjawi

The first model suite for translating between Arabic dialects and to Modern Standard Arabic. It includes Lahjawi-D2D and Lahjawi-D2MSA, showing strong performance in preserving meaning and fluency.

link:🔗https://aclanthology.org/2025.wacl-1.2/

Open Datasets & Benchmark Resources

We are committed to supporting open research by sharing high-quality data.

Misraj Open Data

Over 35 billion Arabic tokens from diverse sources.

View Dataset

Sadeed Tashkeela

A training dataset for automatic diacritization.

View Dataset

SadeedDiac-25

An evaluation benchmark for diacritization quality.

View Dataset

Tarjama-25

An evaluation benchmark for machine translation systems.

View Dataset