taryya/RunningBench-Annotation-Packs

RunningBench Annotation Packs

This repository contains ten self-contained offline packages for human verification of 698 RunningBench multiple-choice video questions.

What is included

• 698 question-specific 480p MP4 clips, split across ten archives.
• Parts 01–09 contain 70 questions each; part 10 contains 68 questions.
• Every archive contains:
  • index.html — offline annotation interface.
  • clips/*.mp4 — one video clip per question.
  • manifest.json — question metadata without reference answers.
  • README.md — package-level instructions.
• index.json records archive sizes and SHA-256 checksums.

Reference answers are intentionally excluded from the distributed packages.

Annotation workflow

1. Download one runningbench_clips_partXX.tar.gz archive.
2. Extract the archive completely. Do not open index.html from inside an archive viewer.
3. Open index.html in Chrome or Edge.
4. Enter the annotator name and review every question.
5. Click 导出 JSON / Export.
6. Return the generated runningbench_partXX_annotation_<annotator>_<date>.json file.

Progress is saved automatically in the browser. Each package uses a separate progress key, so multiple parts can be annotated on the same computer.

Integrity

Use the SHA-256 values in index.json to verify downloaded archives before distribution.

Source dataset

The broader RunningBench project is available at https://huggingface.co/datasets/taryya/RunningBench.