---
pretty_name: "WRBench Human Annotations"
license: apache-2.0
language:
- en
tags:
- video
- video-generation
- world-models
- benchmark
- evaluation
- human-evaluation
- human-feedback
- wrbench
- arxiv:2606.20545
task_categories:
- image-to-video
size_categories:
- 1K<n<10K
configs:
- config_name: pairs
  data_files:
  - split: default
    path: data/pairs.parquet
- config_name: annotation_to_video
  data_files:
  - split: default
    path: data/annotation_to_video.parquet
- config_name: exclusions
  data_files:
  - split: default
    path: data/exclusions.parquet
---

# WRBench Human Annotations

Human **pairwise-comparison** metadata that calibrates the WRBench automated
diagnostic metrics in *Current World Models Lack a Persistent State Core*. Each clean pair is scrubbed to public
fields and mapped to the stable `video_asset_id` values used in
[`WRBench/wrbench-videos`](https://huggingface.co/datasets/WRBench/wrbench-videos).

Release boundary: **1,156 clean pairs** with **2,547 annotator
verdict records**. Maintenance-queue pairs are withheld and listed in the
`exclusions` config so the boundary is auditable.

## Configs & schema

### `pairs`
| Field | Type | Description |
|---|---|---|
| `pair_id` | string | Stable pair identifier |
| `dimension` | string | Evaluation dimension the comparison targets (D1-D6) |
| `video_asset_id_a` / `_b` | string | `sha256:<hash>` of the two compared clips |
| `verdict` | string | Aggregated pairwise outcome (A / B / tie) |
| `verdict_annotators` | list | Anonymized annotator ids (`p1`, `p2`, …) |
| `n_verdict_annotators` | int | Number of annotators contributing a verdict |
| `created_at` | string | Annotation timestamp |

### `annotation_to_video`
`video_asset_id` → `pair_id` index for joining annotations back to clips.

### `exclusions`
Maintenance-queue pairs withheld from the release, with a reason code.

## Load

```python
from datasets import load_dataset

pairs = load_dataset("WRBench/wrbench-human-annotations", "pairs", split="default")
```

## Privacy

Public fields only: `pair_id`, anonymized `verdict_annotators`, timestamps,
`video_asset_id`, and release-safe endpoint metadata. No internal paths, raw
annotator identities, or other PII are included. See `datasheet.md` for the full
data statement.

## Links

- 🎞️ Videos: https://huggingface.co/datasets/WRBench/wrbench-videos
- 📊 Results table: https://huggingface.co/datasets/WRBench/wrbench-results
- 🖼️ Natural-25 prompts: https://huggingface.co/datasets/WRBench/wrbench-natural25
- 🏆 Leaderboard: https://huggingface.co/spaces/WRBench/wrbench-leaderboard
- 🧭 Release collection: https://huggingface.co/collections/WRBench/wrbench-current-world-models-lack-a-persistent-state-core-6a365c717251293c9fc2cc26
- 💻 GitHub: https://github.com/JinPLu/WRBench
- 🌐 Project page: https://jinplu.github.io/WRBench/
- 📄 Paper: https://arxiv.org/abs/2606.20545
- 🤗 HF paper page: https://huggingface.co/papers/2606.20545

## Citation

```bibtex
@article{wrbench2026,
  title   = {Current World Models Lack a Persistent State Core},
  author  = {WRBench Team},
  journal = {arXiv preprint arXiv:2606.20545},
  year    = {2026},
  url     = {https://arxiv.org/abs/2606.20545}
}
```