Cs2 10k A Large Scale Egocentric Counter Strike 2 Dataset

Reka released RekaCS2-10k, a dataset for training “world models”: systems that learn how a first-person view changes in response to what the player does. In its June 2026 announcement, Reka says it rendered first-person video from public professional Counter-Strike 2 match replays, one clip per player per round. Each clip comes with a per-frame file recording the keys pressed, camera-movement changes that serve as proxies for mouse movement, the player's three-dimensional position and camera angles. Because replays are deterministic, the video and the controls that produced it stay aligned frame by frame, which Reka says is the pairing interactive world models need and real-world footage rarely provides. Reka also released the open-source renderer so researchers can produce more clips from other replays. The post presents the dataset as designed for this training and reports no model trained on it.