Reka AI Drops CS2-10k, 600K Gaming Videos to Train Smarter World Models

CS2-10k is a new open dataset from Reka AI that packages over 600,000 first-person Counter-Strike 2 gameplay videos , totaling more than 10,000 hours of footage , with exact per-frame action annotations. Every frame in every clip is paired with the keyboard keys that were pressed, the mouse deltas that moved the camera, and the player's 3D world position at that instant. It is free to download on Hugging Face today under a CC BY-NC 4.0 license.

The data problem world models keep running into

Training interactive world models requires data that is notoriously hard to find: egocentric video sequences with densely aligned action signals , keyboard inputs, camera motion, and ego state , all synchronized to the visual stream. Real sensor data from robotics or AR headsets is expensive to collect and narrow in scope. Synthetic data is cheap but lacks visual richness. Real-world embodied data is costly to collect, while synthetic data often lacks the visual richness or behavioral diversity needed for generalization.

Counter-Strike 2 turns out to be a surprisingly elegant solution to this. Because matches are recorded as deterministic replays, you can reconstruct clean first-person video at any point in a match, extracting the precise control inputs that drove each visual change. The result is a perfectly labeled dataset at essentially zero annotation cost.

What's actually in the dataset

CS2-10k is built from public professional match demos sourced from HLTV. For each demo, Reka renders clean first-person video at 720p, 48fps using the demo replay tool inside CS2, producing one video per player per round. With 10 players per match, that multiplies out fast.

The annotation schema is dense and precise. Each clip ships with a .parquet file where every frame entry contains:

• Actions: a string of active keys , W/A/S/D for movement, J for jump, C for crouch, [ for fire, ] for scope
• Mouse deltas: horizontal and vertical camera movement per frame
• 3D position: position_x, position_y, position_z in game world units
• Camera angles: yaw (−180° to 180°) and pitch (−90° to 90°)

Loading a clip and its annotations is straightforward:

import pandas as pd
# Load annotation file
df = pd.read_parquet("001372c1.parquet")
# Access per-frame data
frames = df["frame_data"].iloc[0]
for frame in frames:
    print(frame["actions"])       # e.g. "WA"
    print(frame["mouse_x_delta"]) # horizontal camera delta
    print(frame["position_x"])    # world X coordinate

A pipeline built to be extended

Alongside the dataset, Reka is releasing cs2-dem-renderer, the open-source pipeline used to produce it. Given a .dem file, it performs a two-pass parse to extract per-player spawn/death intervals and per-frame button inputs, then drives CS2's built-in demo replay system to render first-person video for each player each round. Frames are streamed in real time from CS2's movie output to ffmpeg (VAAPI HEVC), producing .mp4 clips alongside synchronized .parquet annotation files.

The pipeline is written in Go with a C++ server plugin for CS2. Running it requires CS2 installed via Steam, Go 1.21+, and ffmpeg with VAAPI support. A worker mode handles batch processing of entire demo directories with automatic deduplication , the same setup Reka used to produce CS2-10k at scale.

What you can build with it

Recent interactive world models increasingly treat video generation as controllable world simulation rather than passive future prediction. A central requirement is action-conditioned control, where future frames are generated from user inputs , keyboard/mouse commands or high-level instructions. CS2-10k is a direct fit for that paradigm.

Reka specifically calls out four research directions the dataset supports:

• Action-conditioned video generation: Train models to generate the next N frames given the current frame and a keyboard+mouse action sequence. Dense per-frame controls make CS2-10k a natural fit for models like GameNGen, Genie, DIAMOND, and OASIS.
• Egocentric navigation: With 3D player positions and yaw/pitch per frame, the dataset supports learning navigation priors , what does moving forward look like in a confined corridor vs an open site? How does camera control correlate with positional change?
• Long-horizon planning: Full rounds of 60–90 seconds provide significantly longer temporal horizons than most embodied datasets. A model can learn tactical structure: site entry, holds, rotations, and retakes , each as a coherent visual sequence.
• Multi-agent world modeling: All 10 players per match are recorded simultaneously with shared round and map identifiers, making it possible to model how one agent's actions causally affect another's observations.

The real unlock: clean data at game scale

What makes this dataset structurally different from most alternatives is the absence of annotation noise. Each clip is a contiguous segment of a single round from a single player's perspective. There are no mid-round cuts, no editing transitions, and no UI HUD. The camera moves in a physically plausible relationship in the world, and the player weapon is hidden to eliminate sudden visual changes caused by weapon recoil, reloads, and weapon switching.

This matters a lot for world model training. Models that learn to predict the next frame from an action need a tight, consistent action-observation loop. A single dropped frame or UI overlay can corrupt a training sequence. CS2's deterministic replay system eliminates that class of problem entirely.

Interactive video models existed since 2024 (Genie, GameNGen), but only as narrow prototypes. In 2025, breakthroughs in autoregressive diffusion made it possible to take general-purpose video foundation models and make them both interactive and real-time. The bottleneck is now data , specifically, the kind of densely annotated, temporally consistent egocentric data that CS2-10k provides.

Limitations to keep in mind

A few things to factor in before building on this dataset:

• Non-commercial only: the CC BY-NC 4.0 license restricts commercial use. The underlying match demos remain the property of their rights holders.
• Full upload still in progress: the complete 600K+ video dataset is still being uploaded to Hugging Face. A browsable sample subset of 3 full matches is available now in the interactive viewer.
• Single game domain: all footage comes from CS2, a specific visual environment. Generalization to real-world embodied settings or other games will require domain adaptation work.
• Linux-first pipeline: the rendering tool is built for Linux with VAAPI GPU encoding; Windows support requires manual modifications.

The dataset sits in a growing ecosystem of game-derived egocentric data. A comparable effort, EgoCS-400K, was released around the same time and reflects the same community momentum around CS as a research substrate. Reka's contribution stands out for its scale, the quality of its per-frame annotations, and the fact that the rendering pipeline is fully open-sourced , meaning anyone can generate more data from any public CS2 demo file.