TLDR: Researchers introduce Nile-Chat, a new family of large language models specifically for Egyptian Arabic that uniquely supports both Arabic and Latin scripts. Using a novel Mixture-of-Experts approach, Nile-Chat models significantly outperform existing multilingual and Arabic LLMs on new Egyptian evaluation benchmarks, demonstrating a comprehensive methodology for dual-script language adaptation and making all resources publicly available.
A new family of large language models (LLMs) called Nile-Chat has been introduced, specifically designed to understand and generate text in the Egyptian Arabic dialect. What makes Nile-Chat unique is its ability to handle both Arabic and Latin scripts, addressing a significant gap in current LLM development where most models struggle with dual-script languages.
Egyptian Arabic, also known as Masri, is spoken by over 100 million people and is widely understood across the Arab world. A key characteristic of this dialect is its common use in both traditional Arabic script and a Latin-based script often called Arabizi or Franco-Arabic. Existing LLMs for Arabic typically focus on Modern Standard Arabic (MSA) or offer limited support for dialects, and none have been specifically trained to handle the Latin script for a single language across two scripts until now.
The Nile-Chat family includes three main variants: dense models in 4B and 12B parameters, and a Mixture-of-Experts (MoE) model named Nile-Chat-3x4B-A6B. The MoE model uses a novel approach called Branch-Train-MiX (BTX) to combine “script-specialized experts.” This means that different parts of the model are trained specifically on either Arabic-script or Latin-script Egyptian data, and then merged into a single model that can intelligently route text to the appropriate expert. This modular design helps the model adapt without losing performance or efficiency.
The development of Nile-Chat involved a comprehensive training process using newly created dual-script data. This included continual pre-training on large Egyptian Arabic text corpora, such as audio/video transcripts, forum posts, and song lyrics. Following this, the models underwent fine-tuning on a variety of instruction tasks and a final alignment-tuning stage to enhance safety and adjust preferences. Notably, about 25% of the training data was in Latin script, reflecting real-world usage patterns.
To evaluate the performance of Nile-Chat, the researchers introduced new Egyptian evaluation benchmarks. These benchmarks cover a range of tasks, including understanding (like multiple-choice questions) and generative tasks (such as translation and transliteration) in both Arabic and Latin scripts. The results show that Nile-Chat models consistently outperform leading multilingual and Arabic LLMs, including well-known models like LLaMa, Jais, ALLaM, and Qwen2.5, on these Egyptian-specific benchmarks.
For instance, the Nile-Chat-4B model demonstrated superior performance across various Arabic-script benchmarks and significantly outshone competitors on Latin-script benchmarks. The Nile-Chat-12B model further advanced the state-of-the-art, achieving the highest scores on nearly all Arabic-script tasks. The MoE models, Nile-Chat-3x4B-A6B and 2x4B-A6B, strike a balance, excelling particularly in extensive generation and Latin-script processing tasks, and achieving the highest scores across all translation and transliteration tasks and metrics.
Also Read:
- Enhancing Multilingual Reasoning in AI: A New Approach to Language-Consistent Thinking
- Standardizing LLM Evaluation: How Fine-Tuning Improves Model Rankings
This work represents a significant step forward in adapting LLMs for dual-script languages, addressing an often overlooked aspect in modern LLM development. All models, data, and evaluation code are publicly available, encouraging further research in this area. While promising, the researchers acknowledge some limitations, including occasional hallucinations, potential biases in the dataset, and the reliance on Claude for translating English instructions, which might not fully capture Egyptian Arabic nuances. For more technical details, you can refer to the full research paper: Nile-Chat: Egyptian Language Models for Arabic and Latin Scripts.
