LVBench-Relabel / README.md
Sepand's picture
docs: align citation author order with arXiv
04c3aab
|
Raw History Blame Contribute Delete
5.96 kB
metadata
license: cc-by-nc-sa-4.0
language:
  - en
pretty_name: LVBench Relabel
task_categories:
  - video-text-to-text
  - multiple-choice
tags:
  - video
  - long-video
  - benchmark
  - lvbench
size_categories:
  - 1K<n<10K
source_datasets:
  - lmms-lab/LVBench
configs:
  - config_name: relabelled
    default: true
    data_files:
      - split: test
        path: lvbench_relabelled.parquet
  - config_name: original
    data_files:
      - split: test
        path: lvbench_relabelled_original.parquet

LVBench Relabel

These annotations accompany our paper,
What Does Multiple Choice Actually Measure?
Disentangling Video Understanding from Option Exploitation
,
which studies how answer options affect video understanding evaluation.

arXiv · NeurIPS · OpenReview · HuggingFace 🤗

Corrected multiple-choice answer keys for LVBench, a long-video understanding benchmark. Every question in the corpus was independently re-annotated; this repository publishes the resulting keys alongside the originals so the two can be compared directly.

Mirrored from github.com/OnDeckAI/LVBench-Relabel @ 9cd247e. Load with load_dataset("OnDeckAI/LVBench-Relabel", "relabelled") (or "original" for the published keys).

This repository contains annotations only. No video is redistributed here. See Getting the videos.

If you use these annotations, please cite our paper.

What is here

file rows contents
lvbench_relabelled.parquet 1,530 the corrected questions and answer keys
lvbench_relabelled_original.parquet 1,530 the same questions with the published LVBench keys, row-aligned
excluded.csv 19 question ids not present in either file

The two parquets are row-aligned on uid, so a before/after comparison is a column comparison. Against the published keys:

change items share
correct answer repaired 313 20.5%
question stem reworded 137 9.0%
distractors changed 356 23.3%
any of the three 466 30.5%

A repair means the text of the correct option changed. Compare option text, not the answer letter: the annotators both reordered and rewrote option lists, so the letter moves without the answer changing and the answer changes without the letter moving. The letter differs on only 232 rows, which both understates and mismeasures the correction. Case, punctuation and whitespace are normalised before comparison, so a typography fix is not a repair, and reordering alone is not a repair.

Of the 313 repairs, 94 promote text that was already one of the published options and 219 use text the annotators rewrote.

verify_counts.py reproduces every number above from the two parquets:

python verify_counts.py lvbench_relabelled.parquet lvbench_relabelled_original.parquet

19 of LVBench's 1,549 questions were removed by the annotators as unrepairable and appear in neither parquet. Their ids are listed in excluded.csv.

Schema

lvbench_relabelled.parquet:

column description
uid LVBench question id, the join key to the original dataset
key YouTube video id
type video genre (selfmedia, cartoon, live, tv, sport, documentary)
video_info duration, fps and resolution of the source video
question question stem followed by four options, in LVBench's packed (A) … format
answer correct option letter, A to D
question_type capability tags, e.g. ['reasoning'], ['entity recognition']
time_reference evidence window in the source video
source release identifier of the annotation pass
repair_decision repair if the correct option's text changed, keep otherwise
n_options option count; 4 on every row

lvbench_relabelled_original.parquet has the first eight of these columns, with question and answer exactly as published by LVBench.

Every question has exactly four options. repair_decision is derived from the two parquets by the rule above and agrees with them on every row, so it is a sound filter for the repaired items. It replaces the annotators' own classification, which disagreed with the data on 116 rows.

Getting the videos

Videos and the original annotations come from lmms-lab/LVBench, with the videos under video_chunks/. The upstream project is zai-org/LVBench.

Join on uid to align these keys with the videos.

Licence

Released under CC BY-NC-SA 4.0, matching LVBench, whose ShareAlike term requires derivatives to carry the same licence.

LVBench additionally states: "LVBench is only used for academic research. Commercial use in any form is prohibited." That restriction applies to this derivative too.

Neither LVBench nor this repository holds copyright in the source videos, which remain the property of their respective owners.

Citation

These keys were produced for, and are described in:

@inproceedings{dyanatkar2026multiplechoice,
  title     = {What Does Multiple Choice Actually Measure? Disentangling Video
               Understanding from Option-Exploitation},
  author    = {Piercy, Elliott and Chan, Justin and Dyanatkar, Sepand},
  booktitle = {NeurIPS 2026 Workshop: Can We Trust AI Evaluation? Robustness,
               Causality, and Risk in Modern AI Assessment},
  year      = {2026},
  url       = {https://openreview.net/forum?id=qshO3h92ds}
}