1x-technologies/world_model_tokenized_data

1X世界模型压缩挑战数据集是一个用于世界模型压缩挑战的 repository，包含了训练和验证数据集v2.0，大约有100小时的视频数据。数据集包括了模糊处理的视频数据、新的原始视频数据集以及示例脚本。数据集采用了NVIDIA Cosmos Tokenizer离散的DV8×8×8标记，每个分割都被分片存储，包括视频、段索引、状态和元数据文件。此外，还有一个450样本的测试v2.0数据集。数据集中的状态索引定义了人体各关节的角度和速度等信息。

The 1X World Model Compression Challenge Dataset is a repository for the World Model Compression Challenge, containing train and validation datasets v2.0 with approximately 100 hours of video data. The dataset includes blurred video data, a new raw video dataset, and example scripts. The dataset uses NVIDIA Cosmos Tokenizer discrete DV8×8×8 tokens, with each split sharded and stored including video, segment index, state, and metadata files. Additionally, there is a test v2.0 dataset with 450 samples. The state index definition in the dataset defines angles and velocities for various human joints.