shuzhig/LVBench-temporal

LVBench — Temporal Grounding (trimmed)

The temporal grounding subset of LVBench, where each question's video has been pre-cut to just the relevant window. Run inference on a short clip instead of the full (30–110 min) video.

Contents

Path	Description
video_info.meta.jsonl	Metadata + QAs, one JSON object per line (72 lines).
videos/	Pre-cut clips, named <uid>.mp4 (219 clips, one per QA, 3.1 GB).
video_info.meta.jsonl.bak	Original un-cleaned metadata (kept for reference).

72 source videos → 219 questions → 219 trimmed clips. Every uid in the metadata has exactly one matching clip, and every clip is referenced by exactly one QA.

How to use

For each QA, the clip is videos/<uid>.mp4. Feed the clip + the question to a video model and compare its chosen option (A–D) against answer.

import json

for line in open("video_info.meta.jsonl"):
    d = json.loads(line)
    for qa in d["qa"]:
        clip     = f"videos/{qa['uid']}.mp4"   # the trimmed video
        question = qa["question"]              # MCQ text (options A–D inline)
        answer   = qa["answer"]                # correct option letter

To pull the dataset from the Hub:

from huggingface_hub import snapshot_download
path = snapshot_download("shuzhig/LVBench-temporal", repo_type="dataset")

Each top-level record has key (source video id), type (genre), video_info (duration_minutes, fps, resolution), and qa (list of questions).

Each qa object:

• uid — question id; the matching clip is videos/<uid>.mp4.
• question — multiple-choice text. Its last line is an appended note telling the model the clip starts at 00:00 and to treat the whole clip as the relevant window (so it ignores the absolute timestamps in the stem):

  Note: the provided video is a clip trimmed to the original segment MM:SS-MM:SS. The clip itself starts at 00:00, so treat the entire clip as the relevant window and do not rely on the absolute timestamps mentioned above.

• answer — correct option letter (A–D).
• question_type — list of labels; every item includes temporal grounding.
• time_reference — "MM:SS-MM:SS" original-video window the clip was cut from. MM is total minutes and may exceed 59 (e.g. 110:23).

Known data quirk

One QA (key = rSE2YPcv89U) stores uid as the integer 3641 rather than the string "3641"; every other uid is a string. The clip videos/3641.mp4 exists and f-string interpolation works either way, but code that does str-typed set membership or dict lookups on uids should coerce with str(qa["uid"]).

Important for the model

The clips start at 00:00, but the question stems still mention the original timestamps (e.g. "What happens from 01:58-02:46?"). Do not try to seek to those timestamps inside the clip — the clip is that window. Answer from the entire clip.

Distribution

Source genres (72 videos): selfmedia 17 · sport 13 · tv 12 · cartoon 11 · live 10 · documentary 9

Question types (219 QAs, multi-label): temporal grounding 219 · event understanding 92 · entity recognition 62 · reasoning 37 · key information retrieval 24 · summarization 14

Answer balance: D 64 · C 60 · A 54 · B 41

Provenance and license

Derived from LVBench (https://lvbench.github.io/) — questions and answers are the LVBench authors'; the clips here are re-cuts of the original LVBench videos, trimmed to each question's time_reference window, plus an appended note in each question stem. The source videos were collected from public platforms. Intended for non-commercial research evaluation; refer to the upstream LVBench release for the governing terms and cite the original benchmark rather than this derived subset. Rights holders can open a discussion on this repo for removal.