# Mistral 7b Meets Darija 🇩🇿: A Continual Pre-training Journey
Welcome to the **Mistral 7b Meets Darija 🇩🇿** project!
This project aims to develop a **large language model** capable of **understanding** and **generating text** in **Algerian Darija**. We're using **continual pre-training** to teach the model **Arabic** and **French** first, and then **fine-tuning** it on a specifically curated **Algerian Darija dataset**.
#### Technical Approach:
- **Model**: We're primarily using **Mistral-7B**, a powerful and efficient large language model.
- **Training Method**: Continual pre-training on Arabic and French datasets followed by fine-tuning on Algerian Darija data.
- **Tools and Techniques:** Leveraging **PEFT**, **Unsloth Framework**, **RoPE**, and **Flash Attention-2** for optimized training and performance.
The project Details can be found in the This blog post .
## Running the train Script :
before runing the train script, You should install the requirements:
- **For Pytorch 2.1.0** :
```
pip install --upgrade --force-reinstall --no-cache-dir torch==2.1.0 triton \
--index-url
download.pytorch.org
pip install "unsloth[cu118] @ git+
github.com"
pip install "unsloth[cu121] @ git+
github.com"
pip install "unsloth[cu118-ampere] @ git+
github.com"
pip install "unsloth[cu121-ampere] @ git+
github.com"
```
- **For Pytorch 2.1.1:**
```
pip install --upgrade --force-reinstall --no-cache-dir torch==2.1.1 triton \
--index-url
download.pytorch.org
pip install "unsloth[cu118-torch211] @ git+
github.com"
pip install "unsloth[cu121-torch211] @ git+
github.com"
pip install "unsloth[cu118-ampere-torch211] @ git+
github.com"
pip install "unsloth[cu121-ampere-torch211] @ git+
github.com"
```
- **For Pytorch …