hamzabouajila/ar-pusht-image

---
language:
- ar
license: apache-2.0
size_categories:
- 10K<n<100K
task_categories:
- robotics
tags:
- vla
- pusht
- embodiment
- arabic-nlp
pretty_name: Arabic PushT Image Dataset
---

# 🌍 Arabic PushT Image Dataset

## The first Arabic-translated dataset for Vision-Language-Action (VLA) robotics.

This dataset is a localized version of the standard **PushT** image dataset, where every instruction has been translated into high-quality Arabic to enable the training of **localized robotics policies**.

### 🌟 Key Features
- **Native Arabic Instructions:** High-quality translations for complex manipulation tasks.
- **VLA Ready:** Perfectly compatible with `FastVLA` and `OpenVLA` training pipelines.
- **High Performance:** Used to train the [FastVLA Arabic Hero](https://huggingface.co/hamzabouajila/fastvla-arabic-hero) model, achieving 2.3x better accuracy on Arabic commands than base models.

### 📥 Usage
```python
from datasets import load_dataset

dataset = load_dataset("hamzabouajila/ar-pusht-image")
sample = dataset['train'][0]
print(f"Instruction: {sample['instruction']}") # 'دفع الحجر إلى الهدف'
```

### 🧪 Integration
This dataset is designed to be used with the **FastVLA** framework to train robotics policies for $0.48/hr on budget hardware.

---
**Maintained by:** [Hamza Bouajila](mailto:[e-mail]) | [GitHub Repository](https://github.com/BouajilaHamza/fastvla)