Skip to content

Releases: JinPLu/WRBench

WRBench 0.1.3

Choose a tag to compare

@JinPLu JinPLu released this 12 Jul 19:19

WRBench 0.1.3

  • Restores the official Apache-2.0 license and unifies package, CLI, and build version metadata.
  • Adds the canonical OSS verification gate to CI across Python 3.10–3.12.
  • Records the verified rolling D5/D6 correction without changing paper aggregates or video bytes.
  • Synchronizes 59 EasyAnimate rolling rows with the current aggregate Results snapshot.
  • Corrects Spatia and LiveWorld input metadata and distinguishes the frozen 23-model/9,600-row release from the rolling 26-model/11,100-row release.
  • Bundles the 106 exact human-review videos required by all 1,170 recheck endpoints.

Data revision: https://huggingface.co/datasets/WRBench/wrbench-videos/tree/rolling-main-20260713-v3

Verification: 689 tests, clean wheel/sdist build, isolated wheel install, anonymous Hub and Viewer readback, and 106/106 video decode.

SHA256:

  • wheel: f3dd13e786f4e95ea05f242a8d2f8de46f155aa415522d3b02ae78321ff4b667
  • sdist: a469ddebd3119951b36442761accca67509184c9472f5bf9a9800702792bf34c

WRBench v0.1.1 — paper benchmark reproducibility release

Choose a tag to compare

@JinPLu JinPLu released this 11 Jul 20:09
d27e6b2

WRBench v0.1.1 — paper benchmark reproducibility release

This release publishes the immutable paper_main_20260608 provenance contract
for the frozen 23-model, 9,600-output paper benchmark.

It adds the exact local and API-source prompt catalogs, the audited 100-row
Wan2.7 static-camera conditioning-source manifest and preparation flow, the
frozen camera scope, and the HyDRA 77+77 evaluation policy. The public rolling
video dataset remains 11,100 rows.

No benchmark score, generated video, first frame, or annotation was changed.

release_attestation.json records the GitHub and Hugging Face commits, tags,
artifact hashes, CI results, and live Viewer checks.