egocs-400k/dataset

EgoCS-400K is a large research dataset built for training AI models that understand how games are played. It focuses on Counter-Strike 2, the competitive first-person shooter, capturing over 10,000 hours of first-person gameplay video across more than 40,000 match rounds on 13 maps. What makes this dataset unusual is the depth of synchronized data attached to each video clip. Every frame is paired with the exact keyboard and mouse inputs the player pressed at that moment, the game state at tick-level granularity (32 ticks per second, capturing position, weapon, health, and round context), labels for individual actions like firing or reloading, groupings of actions that belong together, and natural-language captions describing what is happening. This combination lets AI researchers train models that connect what a player sees to what they do. The data is organized into a hierarchy of scales: full-round player trajectories, model-ready video clips with stable temporal boundaries, protected action chains (groups of actions that should not be split during training), individual atomic actions, and raw per-tick state traces. A dynamic programming algorithm creates the clip boundaries to avoid cutting sequences mid-action. The dataset is publicly available on Hugging Face, and an interactive viewer on Hugging Face Spaces lets you browse videos alongside their synchronized annotations. A technical report on arXiv covers the construction pipeline in detail. Processing and annotation code are planned for release as the project continues to expand. This is a niche resource aimed at AI researchers working on video understanding, action recognition, or world models for interactive environments. It is not a playable game, a bot, or a tool for casual use.