You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

CMBench merged dataset — 135 cases

This directory merges the 127-case annotation set from 2026-07-27 with the eight-case Chinese Prompt annotation set from 2026-08-03. The original 127 cases retain IDs cmb_0001 through cmb_0127; the eight appended cases are renumbered cmb_0128 through cmb_0135.

Layout

  • videos/cmb_0001.mp4 ... videos/cmb_0135.mp4: self-contained benchmark videos.
  • metadata/benchmark_cases.jsonl: canonical line-delimited metadata.
  • metadata/benchmark_cases.json: the same 135 records as a JSON array.
  • audit/case_provenance.jsonl: mapping from merged IDs to source cases.
  • audit/manual_corrections.json: manual prompt and target corrections.
  • audit/prompt_revisions.jsonl and audit/prompt_revisions.md: reviewed prompt revisions.

Canonical metadata fields

Every case contains:

  • case_id
  • video_file
  • prompt
  • domain_category
  • memory_level
  • object_subtype
  • background_motion
  • target_label
  • target_aliases
  • references
  • discard
  • prior_subset

Optional fields include time_edit (with rev_frame and del_frame) and youtube_url. Each reference contains a normalized ref_id, exact frame, and bbox_xyxy.

The seven time-edited videos are already rendered outputs. The Chinese Prompts in cases cmb_0128 through cmb_0135 are stored as UTF-8 and preserved exactly as saved in the annotation station.

Downloads last month
38