---
title: ArabGuard Egyptian
emoji: 🛡️
colorFrom: green
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
---

# ArabGuard Egyptian

This Docker Space trains the classifier during the image build from the public
[`d12o6aa/ArabGuard-Egyptian-V1`](https://huggingface.co/datasets/d12o6aa/ArabGuard-Egyptian-V1)
dataset, then serves the trained model through Streamlit.

Only the source code, dependency lock, and Docker configuration belong in the
Space repository. Model weights, checkpoints, downloaded datasets, caches, and
dashboard outputs are generated during the Docker build and are ignored by Git.

The first build can take a long time because `xlm-roberta-base` is trained on CPU.
The Dockerfile uses a separate training stage, so downloaded data, Hugging Face
caches, optimizer states, and checkpoints are not copied into the runtime image.
Changing only `app.py` reuses the cached training layer when Docker cache is
available; changing `train_model.py`, `requirements.txt`, or an earlier layer
starts training again.

## Space repository contents

Upload only:

- `.dockerignore`
- `.gitignore`
- `Dockerfile`
- `README.md`
- `app.py`
- `requirements.txt`
- `train_model.py`

Do not upload `.venv`, model weights, checkpoints, dataset files, caches, or
`dashboard_data`.

Before pushing, check the staged file list and sizes:

```bash
git status --short
git ls-files
```

The Space listens on port `7860`, as required by the `app_port` metadata above.