---
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
- config_name: current_review_submissions
  data_files:
  - split: default
    path: data/current_review_submissions.parquet
- config_name: current_review_consensus
  data_files:
  - split: default
    path: data/current_review_consensus.parquet
- config_name: current_review_status
  data_files:
  - split: default
    path: data/current_review_status.parquet
- config_name: review_assets
  data_files:
  - split: default
    path: asset_manifest.parquet
---

# WRBench Human Annotations

This dataset contains the human comparison labels used to validate WRBench's
automatic evaluation metrics.

## Version Update: 2026-07-07

We updated the release after rechecking videos that changed during benchmark
maintenance. The release now includes:

- 1,741 clean comparison rows.
- 4,302 individual human judgments.
- 585 newly rechecked current-benchmark comparisons, each reviewed by three
  annotators.
- Majority-label summaries for the newly rechecked subset.

These counts describe this dated human-label release. Moving leaderboard,
video, and model counts are reported by the results and videos datasets.

## Videos for the current recheck

All 1,170 endpoints in `current_review_consensus` resolve to the exact bytes
shown to annotators. Of these, 897 are in `wrbench-videos`; the remaining 273
endpoints refer to 106 unique files bundled here under `videos/` and indexed by
`asset_manifest.parquet`.

The bundled files are dated human-review evidence. They use the earlier
profile-unsuffixed IDs and are not additional leaderboard videos or aliases for
the 30-degree and 60-degree benchmark profiles. Join on `video_asset_id`, whose
`sha256:` value verifies the file bytes.

## Configs

- `pairs`: clean comparison rows used by the public WRBench release.
- `annotation_to_video`: maps each comparison to the corresponding video assets.
- `current_review_consensus`: majority labels for the 585 newly rechecked
  comparisons.
- `current_review_submissions`: anonymized per-annotator labels for those 585
  comparisons.
- `current_review_status` and `exclusions`: records kept for transparency, but
  not used as clean comparison labels.
- `review_assets`: paths, byte counts, and SHA256 identifiers for the 106
  review-only videos bundled with this dataset.

Annotator identities are anonymized in the public files.

## Links

- Videos: https://huggingface.co/datasets/WRBench/wrbench-videos
- Results: https://huggingface.co/datasets/WRBench/wrbench-results
- Prompts: https://huggingface.co/datasets/WRBench/wrbench-natural25
- Leaderboard: https://huggingface.co/spaces/WRBench/wrbench-leaderboard
- Collection: https://huggingface.co/collections/WRBench/wrbench-current-world-models-lack-a-persistent-state-core
- Paper: https://arxiv.org/abs/2606.20545
- Code: https://github.com/JinPLu/WRBench