datasheet.md

WRBench Human Annotations Datasheet

Motivation

Human pairwise comparisons supporting WRBench diagnostic claims.

Composition

• Clean pairs released: 1156
• Verdict records (sum of n_verdict_annotators): 2547
• Excluded maintenance pairs: documented in exclusions config

Collection

Annotators compared model outputs on controlled pairwise tasks. Identities are pseudonymized (p1, p2, ...).

Recommended uses

Benchmark analysis, calibration of automated metrics, qualitative error studies.

Out of scope

Training annotator models; deanonymization attempts.

Citation

@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}
}