The Video Highlights dataset contains playback links of more than 1,470 complete variety show videos (about 1,200 hours) and files containing extracted features. Highlight Clips includes 18,000 videos with a total length of about 750 hours, and each video is accurately annotated with a timestamp. Researchers can use the videos and annotated clips in the training dataset to design algorithms and train models.

The dataset is divided into three subsets:

• Training set: can be used to design algorithms and train models. It contains video features and label information.

• Validation set: Contains relatively few videos that can be used to validate the trained model. In addition, ground-truth (Meta.json) with timestamps of highlight clips is also provided.

• Test set: Contains another set of videos (without the video highlights timestamp ground-truth). After submitting the predicted timestamp, it can be used to evaluate the model performance and rank.