gnitoahc/vlm-eval-videos

VLM Eval Videos

A video benchmark dataset for evaluating Vision–Language Models (VLMs) on short-form action recognition. Each clip is paired with a fixed question and a ground-truth short-sentence answer, making it suitable for automated VLM inference pipelines and LLM-as-a-judge scoring.

Dataset Details

Description

VLM Eval Videos contains 693 short MP4 video clips drawn from YouTube, organised into five categories. Four categories contain clips of a single repeating action (climbing a ladder, face-planting, falling off a bike, falling off a chair). A fifth category (general_mp4) covers 46 diverse everyday activities sourced from a broader action-recognition corpus.

Each clip is annotated with:

Field	Description
filename	The MP4 filename as stored in the repository
dataset_name	Category directory name (e.g. climbing_ladder)
video_id	YouTube video ID (filename with dataset prefix stripped)
label	Same as dataset_name; used as the ground-truth category tag
question	"What happened? Describe in a single sentence with 2 to 6 words"
answer	Short ground-truth answer sentence (e.g. "The person falls off a bike")

Category Summary

Category	# Clips	Ground-Truth Answer
climbing_ladder	509	The person climbs a ladder
face_planting	46	The person face plants
falling_off_bike	45	The person falls off a bike
falling_off_chair	43	The person falls off a chair
general_mp4	50	Varies (e.g. snowboarding, dancing ballet)
Total	693

The general_mp4 category covers 46 distinct action classes including canoeing, dancing ballet, golf chipping, hula hooping, snowboarding, somersaulting, windsurfing, and more.

Dataset Sources

All video clips originate from publicly available YouTube content. The dataset is intended solely for non-commercial research and evaluation purposes.

Uses

Direct Use

• Offline VLM inference benchmarking: feed sampled frames from each clip to a vision–language model and measure how well it describes the depicted action.
• LLM-as-a-judge scoring: compare model responses against the provided answer field using an LLM evaluator.
• Text-metric evaluation: compute BLEU, ROUGE-L, or CIDEr between model responses and ground-truth answers.

Out-of-Scope Use

• Any commercial application or redistribution of the video content.
• Training or fine-tuning models (the dataset contains no train split).

Dataset Structure

data/
├── climbing_ladder/
│   └── validation/
│       ├── metadata.csv
│       └── climbing_ladder_<video_id>.mp4  (×509)
├── face_planting/
│   └── validation/
│       ├── metadata.csv
│       └── faceplanting_<video_id>.mp4  (×46)
├── falling_off_bike/
│   └── validation/
│       ├── metadata.csv
│       └── falling_off_bike_<video_id>.mp4  (×45)
├── falling_off_chair/
│   └── validation/
│       ├── metadata.csv
│       └── falling_off_chair_<video_id>.mp4  (×43)
└── general_mp4/
    └── validation/
        ├── metadata.csv
        └── <youtube_id>_<start>_<end>.mp4  (×50)

Each metadata.csv follows the same schema and can be loaded independently or combined.

Loading with the datasets library

from datasets import load_dataset

# All categories combined (validation split)
ds = load_dataset("gnitoahc/vlm-eval-videos")

# Single category
ds = load_dataset("gnitoahc/vlm-eval-videos", name="falling_off_bike")

print(ds["validation"][0])
# {
#   "filename": "falling_off_bike_0uf2P7BX1vs.mp4",
#   "dataset_name": "falling_off_bike",
#   "video_id": "0uf2P7BX1vs",
#   "label": "falling_off_bike",
#   "question": "What happened? Describe in a single sentence with 2 to 6 words",
#   "answer": "The person falls off a bike"
# }

Downloading video files

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="gnitoahc/vlm-eval-videos",
    repo_type="dataset",
)

Dataset Creation

Curation Rationale

The dataset was assembled to provide a controlled, reproducible evaluation set for short-form action understanding in VLMs. Four focused action categories (ladder climbing, face-planting, bike falls, chair falls) allow precise measurement of how well a model distinguishes visually similar yet semantically distinct events. The general_mp4 category broadens coverage with diverse everyday activities.

Source Data

Clips were collected from YouTube search results for each action category. general_mp4 clips are 10-second segments drawn from a broader action-recognition corpus, identified by YouTube video ID and temporal offsets encoded in the filename (<id>_<start_sec>_<end_sec>.mp4).

Annotations

Ground-truth answers for the four focused categories are a single canonical sentence per category. general_mp4 answers are the human-annotated action labels from the source corpus.

Considerations for Using the Data

Social Impact

Videos depict real people performing physical activities, including falls and accidents. Users should be mindful of individual privacy and dignity when using or displaying these clips.

Limitations

• The dataset contains only a validation split; it is not suitable for training.
• Category labels are coarse; intra-class variation (camera angle, setting, performer) is not annotated.
• general_mp4 covers 46 action classes with approximately 1 clip per class, making per-class statistics unreliable.
• Video quality and length vary across sources.