VidMsg

VidMsg is a manifest-only benchmark for evaluating implicit message understanding in short, internet-native videos. It contains 400 public YouTube video references paired with 52 fine-grained narrative messages across 9 thematic topics.

VidMsg evaluates whether models can infer what a video communicates, rather than only what it visibly depicts. The repository does not redistribute raw video files. It provides YouTube IDs/links, topic and message labels, metadata, and a CSV-backed evaluation loader.

Files

.
├── README.md
├── LICENSE
├── CODE_LICENSE
├── THIRD_PARTY_VIDEO_NOTICE.md
├── data/
│   └── vidmsg.csv
├── metadata/
│   └── croissant.json
└── code/
    └── dataset.py

Dataset file

data/vidmsg.csv is the canonical manifest. It contains exactly four columns:

Column	Description
video_id	YouTube video identifier.
youtube_link	Public YouTube watch URL.
topic	One of 9 thematic topics.
message	Narrative message conveyed by the video.

Summary:

• Records: 400
• Unique messages: 52
• Topics: 9
• Raw video files included: no
• CSV SHA-256: 99e2455fdcd0a2dfddefba9c37213903d4dbcb4dd8a60a324073fa04a164c35d

Release model and rights

This repository releases annotations, benchmark metadata, and code. It does not release or rehost the source videos.

• data/vidmsg.csv, metadata/croissant.json, README text, labels, and metadata are released under CC BY 4.0.
• code/dataset.py is released under the MIT License.
• Raw YouTube videos are not included and are not licensed by this repository. The corresponding paper states that the clips are under the Standard YouTube License and that annotations are released under CC BY 4.0 with corresponding YouTube URLs.

See THIRD_PARTY_VIDEO_NOTICE.md for the third-party video notice.

Construction summary

VidMsg was built with a message-first pipeline:

1. Define target messages across practical topics.
2. Use LLMs to generate short narrative scenarios that imply each target message without stating it directly.
3. Generate and validate indirect YouTube search queries.
4. Retrieve public YouTube candidates and filter out clips longer than three minutes.
5. Use Amazon Mechanical Turk annotation with three independent workers per candidate.
6. Retain candidates only when a majority of annotators rate message support highly and no annotator marks the clip as rating 1 or too_obvious.

The corresponding paper reports 520 candidates and 400 retained clips.

Evaluation tasks

code/dataset.py supports four evaluation modes:

• text2video: retrieve videos given a target message.
• video2video: retrieve videos that convey the same message as a query video.
• qa_same_topic: five-option multiple-choice QA with distractors from the same topic.
• qa_all_topics: five-option multiple-choice QA with distractors across all topics.

Quick start

pip install pandas torch

import sys
sys.path.insert(0, "code")
from dataset import VidMsgDataset, VidMsgQADataset

retrieval = VidMsgDataset(mode="text2video")
print(len(retrieval))
print(retrieval.get_statistics())

qa = VidMsgQADataset(same_topic=True)
print(qa[0]["formatted_question"])

By default, the loader keeps all 400 manifest records and does not require local .mp4 files.

Limitations and societal impact

VidMsg is a small, fixed, YouTube-derived diagnostic benchmark, not a representative sample of all online videos, cultures, languages, or message types. Implicit message labels are subjective and may vary across viewers and contexts. LLM-assisted story/query generation, YouTube retrieval, platform ranking, and crowdworker judgments may introduce bias.

Potential benefits include improving evaluation of message-level video retrieval, educational media discovery, and analysis of socially meaningful video content. Potential risks include overclaiming model understanding, persuasive targeting, surveillance, profiling, or culturally biased interpretation of latent video messages. VidMsg is intended primarily as a research benchmark.