README.md

---
language:
- en
pretty_name: Wardy Hazard Object Detection Dataset
task_categories:
- object-detection
tags:
- yolo
- object-detection
- hazard-detection
- safety
- wardy
---

# Wardy Hazard Object Detection Dataset

This dataset was prepared for training and evaluating indoor hazard object detection models. It contains images and YOLO-format bounding-box annotations.

## Classes

| ID | Class |
|---:|---|
| 0 | `scissors` |
| 1 | `knife` |
| 2 | `cutter` |
| 3 | `syringe` |

## Versions

| Revision | Description |
|---|---|
| `hazard-objects-v1` | Initial hazard object dataset |
| `hazard-objects-v2` | Updated and extended hazard object dataset |
| `finetuning-v2` | Additional fine-tuning dataset |

Use a revision tag instead of `main` when reproducibility is important.

## Dataset Structure

The `hazard-objects-v1` and `hazard-objects-v2` archives use the following YOLO dataset structure:

```text
dataset/
|-- data.yaml
|-- images/
|   |-- train/
|   |-- val/
|   `-- test/
`-- labels/
    |-- train/
    |-- val/
    `-- test/
```

Each annotation line follows the YOLO format:

```text
class_id x_center y_center width height
```

Coordinates are normalized to values between 0 and 1.

## Download with hf CLI

```bash
hf download chocochip119/hazard \
  --type dataset \
  --revision hazard-objects-v2 \
  --local-dir ./hazard-dataset
```

## Download with Python

```python
from huggingface_hub import hf_hub_download
from zipfile import ZipFile

zip_path = hf_hub_download(
    repo_id="chocochip119/hazard",
    repo_type="dataset",
    filename="dataset.zip",
    revision="hazard-objects-v2",
)

with ZipFile(zip_path) as archive:
    archive.extractall("./hazard_objects_v2")
```

## Training with Ultralytics

After extracting the archive, update the `path` field in `data.yaml` for your environment.

```python
from ultralytics import YOLO

model = YOLO("yolo11n.pt")
model.train(
    data="./hazard_objects_v2/data.yaml",
    epochs=100,
    imgsz=640,
)
```

## Limitations

- Class distributions may be imbalanced.
- Incorrect or missing annotations may exist.
- The dataset may not represent every environment, lighting condition, or camera angle.
- Models trained on this dataset require separate validation before real-world deployment.
- Safety-critical decisions must not rely only on predictions from a model trained with this dataset.

## Source and License

The complete source and redistribution terms of all images have not yet been documented. Verify ownership, consent, privacy requirements, and licensing before using or redistributing this dataset.

## Related Model

- [Wardy M05 Hazard Detector](https://huggingface.co/chocochip119/wardy-m05-hazard-detector)