README.md

metadata

license: other
license_name: research-only-mixed-sources
license_link: LICENSE
viewer: false

BenchCheck / Sextant — data archive (research use only)

Companion data for github.com/Espere-1119-Song/BenchCheck (video-benchmark meta-evaluation, archive freeze 2026-08-24). Restore guide: HANDOFF.md in the GitHub repo.

Research use only. normalized/ mirrors video clips from 189 upstream video benchmarks — most were publicly released by their authors; a subset was re-fetched from YouTube following the benchmarks' official video-ID lists — normalized to ≤480p for evaluation reproducibility. All rights remain with the original content owners and benchmark authors; this mirror exists solely to make the accompanying meta-evaluation reproducible. Do not redistribute outside research use. Takedown: content owners or benchmark authors may open a discussion on this repo and the affected files will be removed promptly.

Layout: loose files + tar volumes

Due to per-file commit limits, most videos ship as tar volumes:

• normalized/… — files committed individually before the switch.
• normalized_tars/normalized_part_XX.tar — the remaining ~135GB; member paths are relative (normalized/<hash>/<file>), so extraction merges into the same tree.
• normalized_tars/tar_manifest.csv — volume,path for every packed file.

Restore the full tree (run at the download root; duplicates are identical, tar wins):

for t in normalized_tars/normalized_part_*.tar; do tar xf "$t"; done

Contents

path	size	what
sextant_data_bundle.tar.gz	472MB	exp/data/ sans frames: SigLIP-2/DINOv3/text embeddings, captions, P1 full per-question results (303,297 rows), per-bench samples (n=300, seed=42), manifests, n0 caches; + exp/watch/logs/ (classifier audit trail)
sextant_archive104.tar.gz	~5.5GB	rescue of GPU-node store: annotations/ (raw benchmark annotation files), p1_raw/, samples/ + samples_av_excluded/ + samples_dropped/ + fix backups, results_local/, logs/ (manifest/failures/dedup), 104-side code mirror, conditions_results.csv
normalized/	136GB	27,233 normalized videos (stage-C output; filenames = <bench>__<safe qid/video id>), the expensive-to-rebuild asset (YouTube portions need fresh cookies)

Rebuild notes

• Frame caches (77GB local + 32GB node) were intentionally dropped: rebuild with exp/pipeline/extract_frames.py from normalized/.
• Zip/tar index caches were dropped: auto-rebuilt on first access.
• Everything else (all result tables) lives in the GitHub repo under exp/results/.