README.md

---
license: mit
task_categories:
  - reinforcement-learning
tags:
  - world-model
  - nes
  - super-mario-bros
  - death-data
size_categories:
  - 10K<n<100K
---

# SMB Death Data (Noisy TAS)

Training data containing **death sequences** for Super Mario Bros world model.

## Dataset Description

- **89,295 frames** from noisy TAS playthroughs
- Noise injected at 10 different points (9%, 18%, 27%... through game)
- 5% noise rate (random button modifications)
- Captures deaths at ALL stages of the game (World 1 through 8)

## Why This Data?

Clean TAS data only shows optimal play - Mario never dies. This dataset adds:
- Goomba collisions → death
- Pit falls → death  
- Koopa collisions → death
- Missed jumps → death
- Failed timing → death

At every stage of the game, not just World 1-1.

## Data Format

Same as smb-worldmodel-data:
```python
data = np.load("frame_000000.npz")
frame = data['frame']   # (224, 256, 3) uint8
action = data['action'] # (8,) float32
```

## Usage

```python
from huggingface_hub import hf_hub_download
import zipfile

path = hf_hub_download(
    repo_id="DylanRiden/smb-death-data",
    filename="smb_death_data.zip",
    repo_type="dataset"
)

with zipfile.ZipFile(path, 'r') as z:
    z.extractall("./death_data")
```

## Combine with Clean Data

For training, mix with clean TAS data:
- [DylanRiden/smb-worldmodel-data](https://huggingface.co/datasets/DylanRiden/smb-worldmodel-data) - 118k clean frames
- This repo - 89k death frames
- **Total: 207k frames**

## License

MIT