---
annotations_creators:
- machine-generated
language:
- en
license:
- mit
multilinguality:
- monolingual
pretty_name: CoCreator Driving Scene
size_categories:
- 1K<n<10K
source_datasets:
- original
task_categories:
- visual-question-answering
- image-to-text
task_ids:
- visual-question-answering
- image-captioning
tags:
- driving
- causal-reasoning
- autonomous-driving
- scene-understanding
configs:
- config_name: default
  data_files:
  - split: train
    path: train-*.parquet
---

# CoCreator Driving Scene

A driving scene causal dataset, automatically constructed by [CoCreator](https://github.com/NIyueeE/cocreator).

Contains 1892 event samples, each consisting of a temporal sequence of driving frames and causal analysis text.

## Loading

```python
from datasets import load_dataset

ds = load_dataset("NIyueeE/cocreator-driving-scene")
```

Each sample contains:

| Field | Type | Description |
|------|------|------|
| `id` | string | Sample ID (0001–1892) |
| `video_frames` | list[bytes] | Preprocessed frame sequence (960×540, JPEG quality 85) |
| `system_prompt` | string | System prompt used for VLM analysis |
| `causal_text` | string | Causal description from two-stage VLM reasoning |
| `simple_text` | string | Baseline description from single-call VLM (no temporal isolation) |

Browse [review.html](review.html) to view a data preview.

## License

MIT