Building a Sudanese Arabic dataset and fine-tuning LLMs to improve representation of this dialect.
# Sudanese Arabic LLM Project
**Towards Representation of Sudanese Arabic Dialect in Large Language Models**
A collaborative initiative to build a high-quality dataset and fine-tune language models that understand and generate Sudanese Arabic.
---
## 🌍 Project Overview
Sudanese Arabic is a widely spoken but underrepresented dialect in the field of Natural Language Processing (NLP). This project aims to:
- Collect and annotate Sudanese Arabic text from diverse sources.
- Create a balanced and labeled corpus.
- Fine-tune Arabic-supportive LLMs (e.g., AraBERT, CAMeL BERT, LLaMA).
- Evaluate model performance on dialect comprehension and generation.
---
## 🤝 How to Contribute
1. **Fork** this repository.
2. **Create a new branch**
`git checkout -b feature/your-task`
3. **Make your changes** and commit.
`git commit -m "Your message"`
4. **Push** your branch and open a **Pull Request**.
5. Use **GitHub Issues** or **Discussions** to coordinate and communicate.
---
## 🧠 Tasks You Can Help With
| Area | Description |
|------|-------------|
| 🗂️ Data Collection | Gather Sudanese Arabic from social media, transcripts, and oral storytelling. |
| 📝 Annotation | Label text using dialect-specific guidelines. |
| 🔧 Script Writing | Automate preprocessing, cleaning, and formatting tasks. |
| 🧪 Model Fine-tuning | Fine-tune LLMs using the Sudanese corpus. |
| 📊 Evaluation | Test model understanding and generation of Sudanese Arabic. |
| 📢 Communication | Help with outreach, documentation, and community involvement. |
---
## 📝 Annotation Guidelines (Summary)
**Goal:** Consistent and culturally-accurate annotation of Sudanese Arabic text.
### 1. Identify Language Variants:
- **Sudanese Arabic** vs. **Modern Standard Arabic (MSA)**.
### 2. Note Regional Vocabulary:
- Tag terms unique to **Khartoum, Darfur, East, North, South Sudan**.
### 3. Normalize Spelling:
- Respect Sudanese usage (e.g., `شنو؟` instead of `ماذا؟`).
### 4. (Optional) Categorize Content:
- Daily Conv …