Datasetpublic

shuzhig/cgbench

CG-Bench (mini) — clue-grounded long-video QA A local mirror of the CG-Bench mini split: 3,000 multiple-choice questions over 1,118 long videos (mean length ~28 min), each question annotated with the clue intervals (second-level time spans) in the video that actually justify the answer. Contents Path Size Description cgbench_mini.json 2.3 MB 3,000 QA items (see schema below) durations.json 38 KB {video_uid: duration_in_seconds} for 1,219 videos… See the full description on the dataset page: https://huggingface.co/datasets/shuzhig/cgbench.

sourceHugging Facecc-by-nc-sa-4.0updated 1mo agoView on Hugging Face

0likes2kdownloads

Dataset Card

CG-Bench (mini) — clue-grounded long-video QA

A local mirror of the CG-Bench mini split: 3,000 multiple-choice questions over 1,118 long videos (mean length ~28 min), each question annotated with the clue intervals (second-level time spans) in the video that actually justify the answer.

Path	Size	Description
`cgbench_mini.json`	2.3 MB	3,000 QA items (see schema below)
`durations.json`	38 KB	`{video_uid: duration_in_seconds}` for 1,219 videos
`subtitles/`	18 MB	519 `.srt` subtitle files, named `{video_uid}.srt`
`subtitles.tar.gz`	6.4 MB	the same subtitles as a single archive
`videos_3fps_480_noaudio/`	117 GB	1,219 `.mp4` videos, resampled to 3 fps / 480p, audio stripped

QA schema

json

{
  "qid": 14,
  "video_uid": "3ZpmN3-2-Zo",
  "question": "In the video, what color is the vehicle driving from the left in the animation?",
  "answer": "White",
  "choices": ["Light gray", "Pale blue", "Gray", "Silver", "White", "Beige", "Ivory"],
  "right_answer": "E",
  "clue_intervals": [[8, 11]],
  "domain": "Art & Culture",
  "sub_category": "Entity Perception",
  "duration": 2741
}

—right_answer is the letter index into choices (A = first choice); answer is the same option as text.
—clue_intervals is a list of [start_sec, end_sec] spans — the minimal evidence a model needs to see. Use it for clue-grounded evaluation, or to build short-context ablations.
—duration is the full video length in seconds (548–6,325 s, mean 1,707 s).

Distribution

Domains (14): Life Record 579 · Electonic/Social Gaming 313 · Music & TV show 305 · Art & Culture 277 · Embodied Expert 252 · Sports & Exercise 238 · Instruction & Knowledge 230 · Special Scenes 205 · GUI 163 · Driving 112 · Animal & Pet 104 · Humor/funny 99 · Security & Health 90 · News 33

Sub-categories (12): Entity Perception 600 · Event Perception 600 · Entity Cognition 300 · Event Cognition 300 · Text Perception 300 · Hallucination 240 · Time Perception 180 · 2D Spatial Perception 180 · Scene Perception 180 · Time Cognition 60 · Text Cognition 45 · Scene Cognition 15

Loading

python

from huggingface_hub import hf_hub_download, snapshot_download
import json

qa = json.load(open(hf_hub_download("shuzhig/cgbench", "cgbench_mini.json", repo_type="dataset")))

# one video
video = hf_hub_download(
    "shuzhig/cgbench",
    f"videos_3fps_480_noaudio/{qa[0]['video_uid']}.mp4",
    repo_type="dataset",
)

# everything except the 117 GB of video
snapshot_download("shuzhig/cgbench", repo_type="dataset",
                  ignore_patterns=["videos_3fps_480_noaudio/*"])

Note that cgbench_mini.json references 1,118 distinct video_uids while the video folder holds 1,219 files, so a few videos have no questions in this split. Subtitles cover 519 videos only — treat them as optional side information, not a guaranteed modality.

Provenance and license

This is a redistribution of the CG-Bench benchmark (Chen et al., CG-Bench: Clue-grounded Question Answering Benchmark for Long Video Understanding), with videos transcoded to 3 fps / 480p without audio to keep the release manageable. Original annotations and videos belong to the CG-Bench authors and the videos' original uploaders; the source videos were collected from public platforms. Use is intended for non-commercial research evaluation. If you are a rights holder and want material removed, open a discussion on this repo.

Please cite the original benchmark rather than this mirror.