JinPLU commited on
Commit
cdf79c8
·
verified ·
1 Parent(s): 3836ce3

Polish cards + leaderboard (paper-accurate D1-D6, gradio_leaderboard)

Browse files
Files changed (2) hide show
  1. README.md +48 -7
  2. datasheet.md +12 -0
README.md CHANGED
@@ -7,12 +7,13 @@ tags:
7
  - video
8
  - benchmark
9
  - human-evaluation
 
10
  - wrbench
11
  - arxiv:2606.20545
12
  task_categories:
13
  - image-to-video
14
  size_categories:
15
- - n<1K
16
  configs:
17
  - config_name: pairs
18
  data_files:
@@ -30,10 +31,33 @@ configs:
30
 
31
  # WRBench Human Annotations
32
 
33
- Scrubbed human pairwise-comparison metadata mapped to public `video_asset_id`
34
- values. Release boundary: **1,156 clean pairs** with **2,547
35
- annotator verdict records**. **1,485 maintenance-queue pairs are excluded**
36
- (see `exclusions` config).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
37
 
38
  ## Load
39
 
@@ -46,8 +70,25 @@ pairs = load_dataset("WRBench/wrbench-human-annotations", "pairs", split="defaul
46
  ## Privacy
47
 
48
  Public fields only: `pair_id`, anonymized `verdict_annotators`, timestamps,
49
- `video_asset_id`, and release-safe endpoint metadata. No internal paths or PII.
 
 
50
 
51
  ## Links
52
 
53
- - Videos: https://huggingface.co/datasets/WRBench/wrbench-videos
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7
  - video
8
  - benchmark
9
  - human-evaluation
10
+ - human-feedback
11
  - wrbench
12
  - arxiv:2606.20545
13
  task_categories:
14
  - image-to-video
15
  size_categories:
16
+ - 1K<n<10K
17
  configs:
18
  - config_name: pairs
19
  data_files:
 
31
 
32
  # WRBench Human Annotations
33
 
34
+ Human **pairwise-comparison** metadata that calibrates the WRBench automated
35
+ diagnostic metrics in *Current World Models Lack a Persistent State Core*. Each clean pair is scrubbed to public
36
+ fields and mapped to the stable `video_asset_id` values used in
37
+ [`WRBench/wrbench-videos`](https://huggingface.co/datasets/WRBench/wrbench-videos).
38
+
39
+ Release boundary: **1,156 clean pairs** with **2,547 annotator
40
+ verdict records**. Maintenance-queue pairs are withheld and listed in the
41
+ `exclusions` config so the boundary is auditable.
42
+
43
+ ## Configs & schema
44
+
45
+ ### `pairs`
46
+ | Field | Type | Description |
47
+ |---|---|---|
48
+ | `pair_id` | string | Stable pair identifier |
49
+ | `dimension` | string | Evaluation dimension the comparison targets (D1-D6) |
50
+ | `video_asset_id_a` / `_b` | string | `sha256:<hash>` of the two compared clips |
51
+ | `verdict` | string | Aggregated pairwise outcome (A / B / tie) |
52
+ | `verdict_annotators` | list | Anonymized annotator ids (`p1`, `p2`, …) |
53
+ | `n_verdict_annotators` | int | Number of annotators contributing a verdict |
54
+ | `created_at` | string | Annotation timestamp |
55
+
56
+ ### `annotation_to_video`
57
+ `video_asset_id` → `pair_id` index for joining annotations back to clips.
58
+
59
+ ### `exclusions`
60
+ Maintenance-queue pairs withheld from the release, with a reason code.
61
 
62
  ## Load
63
 
 
70
  ## Privacy
71
 
72
  Public fields only: `pair_id`, anonymized `verdict_annotators`, timestamps,
73
+ `video_asset_id`, and release-safe endpoint metadata. No internal paths, raw
74
+ annotator identities, or other PII are included. See `datasheet.md` for the full
75
+ data statement.
76
 
77
  ## Links
78
 
79
+ - 🎞️ Videos: https://huggingface.co/datasets/WRBench/wrbench-videos
80
+ - 📊 Results table: https://huggingface.co/datasets/WRBench/wrbench-results
81
+ - 🏆 Leaderboard: https://huggingface.co/spaces/WRBench/wrbench-leaderboard
82
+ - 📄 Paper: https://arxiv.org/abs/2606.20545
83
+
84
+ ## Citation
85
+
86
+ ```bibtex
87
+ @article{wrbench2026,
88
+ title = {Current World Models Lack a Persistent State Core},
89
+ author = {WRBench Team},
90
+ journal = {arXiv preprint arXiv:2606.20545},
91
+ year = {2026},
92
+ url = {https://arxiv.org/abs/2606.20545}
93
+ }
94
+ ```
datasheet.md CHANGED
@@ -22,3 +22,15 @@ Benchmark analysis, calibration of automated metrics, qualitative error studies.
22
  ## Out of scope
23
 
24
  Training annotator models; deanonymization attempts.
 
 
 
 
 
 
 
 
 
 
 
 
 
22
  ## Out of scope
23
 
24
  Training annotator models; deanonymization attempts.
25
+
26
+ ## Citation
27
+
28
+ ```bibtex
29
+ @article{wrbench2026,
30
+ title = {Current World Models Lack a Persistent State Core},
31
+ author = {WRBench Team},
32
+ journal = {arXiv preprint arXiv:2606.20545},
33
+ year = {2026},
34
+ url = {https://arxiv.org/abs/2606.20545}
35
+ }
36
+ ```