pufanyi/VBVR-Bench

---
language:
- en
license: apache-2.0
task_categories:
- visual-question-answering
- video-classification
tags:
- video
- reasoning
- benchmark
- i2v
pretty_name: VBVR-Bench
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: in_domain
path: data/in_domain-*
- split: out_of_domain
path: data/out_of_domain-*
---

# VBVR-Bench

Re-hosted copy of [Video-Reason/VBVR-Bench-Data](https://huggingface.co/datasets/Video-Reason/VBVR-Bench-Data),
converted to standard HuggingFace parquet format.

## Splits
- **`in_domain`**: 50 tasks x 5 samples = 250 entries (tasks overlap with the VBVR training set).
- **`out_of_domain`**: 50 tasks x 5 samples = 250 entries (held-out reasoning tasks).

## Schema

| field | type | notes |
|---|---|---|
| `task_name` | string | e.g. `G-13_grid_number_sequence_data-generator` |
| `video_idx` | string | zero-padded sample id (`00000`..`00004`) |
| `domain` | string | duplicates split name; convenient for filtering |
| `prompt` | string | task description fed to the I2V model |
| `first_frame` | Image (PNG) | I2V condition frame |
| `final_frame` | Image (PNG) | expected final frame |
| `ground_truth_video` | binary (MP4) | reference video — decode with decord / PyAV |

## Quick load

```python
from datasets import load_dataset

ds = load_dataset("pufanyi/VBVR-Bench", split="in_domain")
sample = ds[0]
sample["first_frame"] # PIL.Image
sample["prompt"] # str
sample["ground_truth_video"] # raw MP4 bytes

# Decode the video with decord
import decord, io
vr = decord.VideoReader(io.BytesIO(sample["ground_truth_video"]))
```

## Links
- Upstream dataset: [Video-Reason/VBVR-Bench-Data](https://huggingface.co/datasets/Video-Reason/VBVR-Bench-Data)
- Evaluation kit: [Video-Reason/VBVR-EvalKit](https://github.com/Video-Reason/VBVR-EvalKit)
- Project page: [video-reason.com](https://video-reason.com/)