VideoMemory-Bench

VideoMemory-Bench is a multiple-choice video question answering benchmark designed to evaluate video memory, long-range temporal understanding, cross-modal association, and reasoning over video content. The released benchmark split contains 5,360 QA pairs over 3,587 unique videos. Each QA item is attached to one video, one normalized level, one normalized category code, and one answer label.

This dataset card describes the Hugging Face release format. The public release renames videos to stable anonymous file names such as vmb_000001.mp4 and renumbers QA pairs in a VideoMME-style format such as 000001-001.

Dataset Summary

• Task: multiple-choice video question answering.
• Modality: video + text question + text options.
• Split: test.
• Number of QA pairs: 5,360.
• Number of unique videos: 3,587.
• Total unique-video duration: 418.67 hours.
• Question language: English.
• Answer format: one option letter, e.g. A, B, C, D, or E.
• Option count: variable; each question has 2 to 5 options.
• Video naming: public videos are renamed as videos/vmb_000001.mp4, videos/vmb_000002.mp4, etc.
• QA naming: public QA IDs use {video_id}-{question_index}, e.g. 000001-001.

Repository Structure

VideoMemory-Bench/
  README.md
  data/
    test.jsonl
  videos/
    vmb_000001.mp4
    vmb_000002.mp4
    ...
  metadata/
    video_map.jsonl
    category_schema.json
    checksums.sha256

Main Annotation File

data/test.jsonl contains one JSON object per QA pair.

{
  "video_id": "000001",
  "video": "videos/vmb_000001.mp4",
  "duration": "medium",
  "duration_sec": 516.27,
  "source_dataset": "LongVideoBench",
  "original_video_id": "lvb_8hhcFRoR0mw",
  "question_id": "000001-001",
  "original_question_id": "lvb_15",
  "level": "L2",
  "level_name": "Motion & Cross-Modal Association",
  "category_code": "2.1",
  "task_type": "Action & State Recognition",
  "question": "On a brown floor, a woman with black hair, wearing a white top and a green skirt, is kneeling on a white mat. On the leftmost side of the white table in front of her, there's a metal pot. What is this woman doing?",
  "options": [
    "A. Making dumplings",
    "B. Watching TV",
    "C. Listening to music",
    "D. Drinking water",
    "E. Stir-frying vegetables"
  ],
  "answer": "A"
}

Field Definitions

Field	Type	Description
video_id	string	Public six-digit video ID assigned by first appearance in the release annotation file.
video	string	Relative path to the renamed video file in this repository.
duration	string	Duration bucket: short, medium, or long.
duration_sec	float	Video duration in seconds.
source_dataset	string	Source collection name before release normalization.
original_video_id	string	Original internal video identifier retained for traceability.
question_id	string	Public QA ID in {video_id}-{index} format.
original_question_id	string	Original internal QA identifier retained for traceability.
level	string	Normalized benchmark level: L1 to L5.
level_name	string	Human-readable level name.
category_code	string	Normalized numeric category code.
task_type	string	Human-readable category name.
question	string	Multiple-choice question.
options	list[string]	Candidate answers formatted as A. ..., B. ..., etc.
answer	string	Correct option letter.

metadata/video_map.jsonl maps each public video ID to its release filename and original identifier. Local absolute paths used during preprocessing are not part of the public release.

metadata/category_schema.json stores the normalized level and category taxonomy used by data/test.jsonl.

Duration Buckets

Duration buckets follow the same broad short/medium/long convention used in long-video evaluation:

Bucket	Rule
short	duration_sec < 180
medium	180 <= duration_sec < 900
long	duration_sec >= 900

Data Statistics

Overall

Metric	Value
QA pairs	5,360
Unique videos	3,587
Total unique-video duration	418.67 hours
Minimum video duration	1.25 seconds
Median video duration	26.01 seconds
Average video duration	420.19 seconds
Maximum video duration	7,331.29 seconds

Source Distribution

Source dataset	QA pairs	Unique videos
LongVideoBench	203	101
MVBench	2,731	2,539
VSI-Super-Count	358	37
VSI-Super-Recall	12	12
Video-MME	1,983	880
VideoMem-Bench	73	18

Duration Distribution

Duration bucket	QA pairs	Unique videos
short	3,176	2,830
medium	467	265
long	1,717	492

Level Distribution

Level	Level name	QA pairs
L1	Basic Visual Perception	393
L2	Motion & Cross-Modal Association	393
L3	Long-range Video Memory	3,306
L4	Reasoning & Commonsense	1,186
L5	Robustness & Negative Memory	82

Option and Answer Distribution

Number of options	QA pairs
2	64
3	896
4	4,303
5	97

Answer label	QA pairs
A	1,620
B	1,670
C	1,365
D	685
E	20

Category Taxonomy

All public records use normalized numeric category codes. The original working files contained mixed category labels such as L3, L4, l3_counting, and CROSS; these are normalized before release.

Level	Level name	Category code	Task type	QA pairs
L1	Basic Visual Perception	1.1	Object & Attribute Recognition	293
L1	Basic Visual Perception	1.2	Video OCR / Text Spotting	100
L2	Motion & Cross-Modal Association	2.1	Action & State Recognition	309
L2	Motion & Cross-Modal Association	2.2	Speed & Trajectory	81
L2	Motion & Cross-Modal Association	2.3	Cross-Modal Associative Memory	3
L3	Long-range Video Memory	3.1.1	Counting	571
L3	Long-range Video Memory	3.1.2	Entity State Evolution / State Tracking	550
L3	Long-range Video Memory	3.2.1	Scene & Synopsis	784
L3	Long-range Video Memory	3.2.2	Visual Needle-in-a-Haystack (NIAH)	123
L3	Long-range Video Memory	3.3.1	Action Sequencing & Temporal Localization	770
L3	Long-range Video Memory	3.3.2	Duration Estimation	128
L3	Long-range Video Memory	3.4.1	Dynamic Spatial Relationships	82
L3	Long-range Video Memory	3.4.2	Egocentric Navigation	225
L3	Long-range Video Memory	3.4.3	3D Layout Inference	73
L4	Reasoning & Commonsense	4.1.1	Explanatory / Why-QA	238
L4	Reasoning & Commonsense	4.1.2	Physical Commonsense	85
L4	Reasoning & Commonsense	4.1.3	Feasibility	67
L4	Reasoning & Commonsense	4.2.1	Intent & Belief Tracking (Theory of Mind)	119
L4	Reasoning & Commonsense	4.2.2	Social Norms & Script Inference	69
L4	Reasoning & Commonsense	4.2.3	Pragmatic Inference & Deception Detection	45
L4	Reasoning & Commonsense	4.3.1	Counterfactual	236
L4	Reasoning & Commonsense	4.3.2	Hypothetical Scenarios	55
L4	Reasoning & Commonsense	4.3.3	Anticipation / Prediction	272
L5	Robustness & Negative Memory	5.1	Negative Memory & Hallucination Detection	82

Evaluation Protocol

Models should answer each question with exactly one option letter. The default metric is exact-match accuracy over normalized answer labels.

Recommended prompt format:

Question: {question}
Options:
{options}
Answer with the option letter only.

Prediction normalization should strip whitespace and punctuation, then take the first valid label among A, B, C, D, and E. A prediction is correct when the normalized label equals answer.

Loading the Dataset

Using datasets:

from datasets import load_dataset

dataset = load_dataset("YOUR_ORG/VideoMemory-Bench", split="test")
print(dataset[0])

Using plain Python:

import json
from pathlib import Path

rows = [
    json.loads(line)
    for line in Path("data/test.jsonl").read_text().splitlines()
]

Each row contains a relative video path. Join it with the dataset repository root to locate the corresponding video file.

Preprocessing and Normalization

The release uses the 5,360-QA benchmark version. During release packaging:

1. Videos are renamed to vmb_000001.mp4, vmb_000002.mp4, etc.
2. QA pairs are renumbered by public video ID and within-video question index.
3. Mixed category codes are normalized to numeric category codes.
4. The normalized level and level_name fields are derived from category_code.
5. Options are formatted consistently as A. ..., B. ..., etc.
6. Answers are normalized to single option letters.
7. Local machine paths are removed from public annotation files.

The source annotation file for this packaged split has SHA-256:

ee0b4d6d4dcbac8ef0a3c7df4286657983c2eb42c415fe6b7ef9329a3eea1217

Intended Use

VideoMemory-Bench is intended for research on video-language models, long-video understanding, video memory, temporal grounding, multimodal reasoning, and robustness to negative or hallucinated memory claims.

Appropriate uses include:

• Evaluating multiple-choice video QA accuracy.
• Comparing models across short, medium, and long videos.
• Analyzing performance by level and task category.
• Studying failures in long-range temporal reasoning and memory retrieval.

Limitations

• The benchmark is multiple-choice; it does not measure open-ended generation quality directly.
• Option counts vary from 2 to 5, so evaluation code should not assume exactly four options.
• Some videos originate from existing public video QA resources; users should check the upstream dataset terms before redistribution or commercial use.
• The benchmark can contain visual scenes involving people, public online videos, subtitles, and other real-world content.
• Answer distributions are not perfectly balanced, especially for option E.
• Public video renaming improves release consistency but does not anonymize the semantic content of the videos.

License and Terms

TODO before public release: specify the final dataset license and any required upstream attribution or usage constraints.

Because the benchmark aggregates or derives examples from multiple source collections, downstream users are responsible for following the applicable terms for the videos and annotations.

Citation

If you use VideoMemory-Bench, please cite the dataset paper or technical report.

@misc{videomemorybench,
  title        = {VideoMemory-Bench},
  author       = {TODO},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  url          = {TODO}
}

Contact

For questions about the benchmark, data packaging, or evaluation protocol, please contact:

TODO: maintainer name / email / project page