README.md

---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: domain
    dtype: string
  - name: task_type
    dtype: string
  - name: prompt
    dtype: string
  - name: image
    dtype: image
  - name: reference_frames
    sequence: image
  - name: reference_text
    sequence: string
  - name: protocol
    sequence: string
configs:
- config_name: default
  data_files:
  - split: train
    path: dataset.parquet
license: mit
language:
- en
---

# Beyond the Last Frame: Process-aware Evaluation for Generative Video Reasoning


## 👀 About VIPER

- Overview: Process-aware evaluation for Generative Video Reasoning tasks.
- Statistics: 309 carefully curated samples spanning 6 distinct domains (i.e., temporal, structural, symbolic, spatial, physics and planning reasoning).
- New Metric: Process-outcome Consistency (POC@r). POC@r evaluate video correctness at both process- and outcome-level, with multiple frames uniformly sampled from the whole video at rate r, instead of the last frame only.

<p align="center">
    <img src="https://huggingface.co/datasets/Monosail/VIPER/resolve/main/overview.png" width="85%"> <br>
</p>

## Dataset Statistics

### Domain Distribution

| Domain | Total Samples | Task Types |
|--------|---------------|------------|
| Physics | 32 | experiment, game |
| Planning | 44 | navigation, obj_manipulation |
| Spatial | 60 | block_rotate, dice, image_restore |
| Structural | 70 | chess, maze, sudoku, ttt |
| Symbolic | 60 | knowledge, math, multimodal |
| Temporal | 43 | obj_move, zoom |

## 📦 Dataset Usage

### Download

```python
from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("Monosail/VIPER")

```

### Data Fields

- `id`: Unique identifier for the sample
- `domain`: The reasoning domain (Physics, Planning, Spatial, Structural, Symbolic, Temporal)
- `task_type`: Specific task category within the domain
- `prompt`: Text prompt describing the task
- `image`: The input image
- `reference_frames`: Ground-truth image frames
- `reference_texts`: Ground-truth text descriptions
- `protocol`: Process-level task constraints



## 📝 Citation

If you find our benchmark useful, please consider citing us:

```bibtex
@article{li2026viper,
  title={Beyond the Last Frame: Process-aware Evaluation for Generative Video Reasoning},
  author={Li, Yifan and Gu, Yukai and Min, Yingqian and Liu, Zikang and Du, Yifan and Zhou, Kun and Yang, Min and Zhao, Wayne Xin and Qiu, Minghui},
  journal={arXiv preprint arXiv:2512.24952},
  year={2025}
}
```