Datasets:

MINT-SDSU
/

CLIP-CC

language:
  - en
license: cc-by-4.0
size_categories:
  - n<1K
task_categories:
  - video-text-to-text
  - summarization
pretty_name: CLIP-CC
tags:
  - video
  - video-captioning
  - video-understanding
  - long-form-video
  - movie
  - benchmark

📚 CLIP-CC Dataset (Movie Clips Edition)

Paper | arXiv | Project Page | Benchmark Code (CLIP-CC-Bench) | Dataset Repo (CLIP-CC)

CLIP-CC is a curated dataset for long-form video description: 200 movie clips sourced from YouTube, each about 90 seconds long (~5 hours in total) and drawn from more than 140 films spanning 1959–2024, each paired with one human-written English reference description averaging 402 ± 208 words. The references were written by four graduate-student annotators over roughly 150 hours, transcribed from live narration and then given grammar-only cleanup, and they deliberately contain no proper nouns, so they describe what is visible on screen rather than naming films, characters, or actors. CLIP-CC is the dataset; CLIP-CC-Bench is the evaluation framework and leaderboard built on it, which scores 17 video-language models with 5 embedding judges.

Dataset Summary

🎥 200 movie clips from well-known films
📺 Each clip is referenced via a YouTube link
📝 Each clip includes a human-written summary describing the scene
💾 All metadata is stored in a .jsonl file (metadata.jsonl)

Use Cases

Video summarization and caption generation
Vision-Language alignment
Video QA and downstream educational AI tasks
Fine-tuning multimodal models (e.g., Flamingo, Video-BERT, LLaVA)

Dataset Structure

Each entry has:

Field	Description
`id`	Unique ID, e.g., `001`
`file_link`	YouTube link to the movie clip
`summary`	Human-written summary of video content

Example

{
  "id": "001",
  "file_link": "https://www.youtube.com/watch?v=-yyoLJuNIJU",
  "summary": "The video begins with a bespectacled man in a white shirt driving a car while speaking. The camera cuts to a black sedan driving down the street, with someone in the passenger seat firing a weapon out of the window. ... (truncated; the full reference for this clip is 648 words)"
}

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("MINT-SDSU/CLIP-CC")

# Access a sample entry
print(dataset["train"][0])

Sample Output


{
  "id": "001",
 "file_link": "https://www.youtube.com/watch?v=-yyoLJuNIJU",
 "summary": "The video begins with a bespectacled man in a white shirt driving a car while speaking. The camera cuts to a black sedan driving down the street, with someone in the passenger seat firing a weapon out of the window. The scene transitions back to the bespectacled man, who now appears increasingly restless. A man in a yellow suit occupies the passenger seat, and there seems to be another individual in the backseat. The focus shifts to the street, where a turquoise sedan is being pursued by the black sedan. A passenger in the black sedan continues firing. The camera returns to the bespectacled man, who is driving anxiously, and pans to the man in the yellow suit as he opens the glove box and retrieves an object. It is revealed to be a gun, which he holds in his hand. The bespectacled man, appearing frustrated, speaks while driving. The two men argue loudly as the camera alternates between them. The turquoise sedan is shown speeding down the street, with pedestrians taking cover. Meanwhile, the black sedan continues its pursuit. The camera briefly focuses on a man wearing a hat in the backseat of the black sedan, who is yelling. It then pans back to the man in the yellow suit, now seen with his feet up on the dashboard, as the vehicles race through the streets. The focus shifts again to the turquoise sedan, capturing both its interior and exterior. Inside the car, the bespectacled man and the man in the yellow suit switch places, with the latter now taking the driver’s seat. The windshield wipers move back and forth as the turquoise sedan continues down the road. The bespectacled man, now in the passenger seat, fires a weapon toward the pursuing black sedan. In the backseat, the man with the hat is shown screaming. The camera returns to the black sedan, where the passenger leans out of the window and continues firing. Meanwhile, the man in the yellow suit, visibly anxious, drives the turquoise sedan while speaking. The street view changes, showing a blue sedan on the left of the turquoise sedan. A side angle shot captures the speeding vehicles, followed by an aerial perspective of the chase. The focus shifts to a black sedan parked in the middle of the street. Inside the turquoise sedan, the bespectacled man and the man in the yellow suit are seen arguing. The scene transitions to a video game-like mode, displaying speed annotations (in mph) and a score in the top-right corner. The video game perspective shows the street, including pedestrians on the curb. The camera then captures the turquoise sedan from the perspective of a man in sunglasses inside a parked car. This man, holding a gun, begins firing while another individual occupies the driver’s seat. The turquoise sedan swerves erratically, driving on the wrong side of the road and drifting uncontrollably. Inside, the man with the hat lies in the backseat, screaming. The focus shifts to two black sedans. One of them speeds past a black sedan parked in the middle of the street, running a red light in the process. The chase intensifies as the black sedan pulls up parallel to the turquoise sedan on its left. Inside the turquoise sedan, the man in the yellow suit turns the steering wheel sharply toward the black sedan.  The camera captures a red SUV parked on the right side of the street, with a man in a white shirt standing nearby on the curb. The black sedan crashes into the red SUV, causing an explosion and shattering glass. The bespectacled man, now in the passenger seat of the turquoise sedan, looks back at the scene, while the man in the yellow suit continues driving. The video concludes with a shot of the turquoise sedan speeding down the street as the black sedan, engulfed in flames, is shown mid-air."
 }

Disclaimers

The CLIP-CC dataset contains references (via YouTube links) to publicly available movie clips. No video files are hosted or redistributed through this dataset—only links to existing content on YouTube are included.
All rights to the original video content remain with their respective copyright holders.
The human-written summaries and metadata in this repository are released under CC BY 4.0; the linked source videos are not covered by that license.
The dataset is intended for research and educational purposes only, such as evaluating video summarization models, vision-language alignment, or building multimodal benchmarks.
Users of this dataset are responsible for ensuring their usage complies with local laws and platform terms of service, including YouTube's Terms of Service.
Record 034 is a podcast segment rather than a movie clip. It is retained unchanged because it was part of the released dataset and of the evaluation reported in the paper; removing it would break exact reproduction of the published results.
If you're a copyright holder and believe any content should not be referenced in this dataset, please reach out through the appropriate platform.

Citation

@misc{ali2026clipccbench,
  title         = {{CLIP-CC-Bench}: Evaluating Paragraph-Level Video Descriptions in Video--Language Models},
  author        = {Ali, Mukhtiar and Dubey, Harsh and Mishra, Sugam and Pack, Chulwoo},
  year          = {2026},
  eprint        = {2608.04302},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  note          = {Presented at the 2nd Workshop on Evaluation for Multimodal Generation (EvalMG), ACM SIGIR 2026},
  url           = {https://arxiv.org/abs/2608.04302}
}