Video Training Dataset

该数据集是一个精心筛选的视频训练集，涵盖了包括驾驶、人体运动、自然动态等多个物理人工智能应用类别。该数据集包含了不同长度和分辨率的视频（从720p到4k），并经过多步骤的精选过程，以提高数据质量，更好地服务于模型训练。该数据集规模约为2000万小时的原始视频，其中约1亿个视频片段用于预训练，1000万个视频片段用于微调。其任务是训练物理人工智能模型。

This dataset is a carefully curated video training set covering multiple physical artificial intelligence application categories, including driving scenarios, human motion, natural dynamics, and more. It contains videos of varying lengths and resolutions ranging from 720p to 4K, and has undergone a multi-step screening process to enhance data quality, thereby better supporting model training. The dataset totals approximately 20 million hours of raw video, with around 100 million video clips allocated for pre-training and 10 million clips reserved for fine-tuning. Its core purpose is to train physical artificial intelligence models.