obaydata/world-model-gameplay-recording

World Model Gameplay Recording

Action-conditioned gameplay video dataset for world model training. Contains synchronized high-resolution gameplay recordings with frame-accurate input action logs (gamepad, keyboard, mouse) from multiple AAA game titles.

Games Included

#	Game	Session ID	Duration	Video Format	Video Size	Input Type
1	Game Session 1	fwa0NekU	~5 min	MKV	582 MB	Gamepad + Keyboard
2	The Legend of Zelda: Tears of the Kingdom	g4qz1DLq	~15 min	MP4 (1080p)	1.29 GB	Gamepad (axis + buttons)
3	The Witcher 3: Wild Hunt	g50N33nG	~13 min	MP4 (1080p)	1.07 GB	Keyboard + Mouse

Total: ~33 minutes of gameplay, ~2.94 GB of video

Dataset Structure

fwa0NekU/                              # Game Session 1
├── raw_videos/
│   └── 2026-02-11 11-17-33.mkv        # Screen recording (~582 MB)
└── raw_meta_data/
    └── 2026-02-11_11-17-33/
        ├── timeline.txt                 # Session start/stop
        ├── gamepad_axis_0.txt           # Analog stick (5103 events)
        ├── gamepad_button_0.txt         # Button press/release (2168 events)
        ├── key_0.txt                    # Keyboard (8 events)
        └── mouse_wheel_0.txt            # Mouse wheel

g4qz1DLq/                              # Zelda: Tears of the Kingdom
├── raw_videos/
│   └── 2026-03-10_23-16-30.mp4        # Gameplay recording (1.29 GB, 1080p)
└── raw_meta_data/
    └── 2026-03-10_23-16-30/
        ├── timeline.txt                 # Session start/stop
        ├── gamepad_axis_0.txt           # Analog stick (continuous)
        ├── gamepad_button_0.txt         # Button events
        ├── mouse_move_0.txt             # Mouse movement
        ├── mouse_pressed_0.txt          # Mouse clicks
        └── mouse_wheel_0.txt            # Mouse wheel

g50N33nG/                              # The Witcher 3: Wild Hunt
├── raw_videos/
│   └── 2026-03-11_10-31-46.mp4        # Gameplay recording (1.07 GB, 1080p)
└── raw_meta_data/
    └── 2026-03-11_10-31-46/
        ├── timeline.txt                 # Session start/stop
        ├── key_0.txt                    # Keyboard events
        ├── mouse_pressed_0.txt          # Mouse button events
        └── mouse_wheel_0.txt            # Mouse wheel events

Data Format

Video

• Format: MKV / MP4
• Resolution: Up to 1080p
• Content: Full-screen gameplay recordings

Action Logs (plain text, one event per line)

timeline.txt — Session boundaries:

2026-03-10 23:16:30.389
2026-03-10 23:16:30.474: obs_recording_started
2026-03-10 23:31:32.958

gamepad_axis_0.txt — Analog stick positions:

2026-03-10 23:16:31.709: axis_1,d,-0.105    # Left stick Y-axis
2026-03-10 23:16:31.715: axis_0,d,0.103     # Left stick X-axis

Format: <timestamp>: <axis_id>,<direction>,<value>

gamepad_button_0.txt — Button presses:

2026-03-10 23:16:33.996: button_1,d          # Button pressed (d=down)
2026-03-10 23:16:34.511: button_1,u          # Button released (u=up)

Format: <timestamp>: <button_id>,<d|u>

key_0.txt — Keyboard events:

2026-03-11 10:31:50.017: w,KEY_DOWN          # W key pressed
2026-03-11 10:31:50.345: w,KEY_UP            # W key released

Format: <timestamp>: <key>,<KEY_DOWN|KEY_UP>

mouse_pressed_0.txt — Mouse button events:

2026-03-11 10:31:48.948: left,d              # Left button pressed
2026-03-11 10:31:49.061: left,u              # Left button released

All actions are timestamped to millisecond precision for frame-accurate alignment with the video stream.

Use Cases

• World Model Pre-training: Learn environment dynamics from video + action pairs
• Action-Conditioned Video Prediction: Predict next frames given current frame + action
• Game Environment Simulation: Train neural game engines
• Game AI / Agent Training: Offline RL and imitation learning from human gameplay
• Video Understanding: Temporal reasoning over complex 3D game environments

Collection Method

• Gameplay recorded using OBS Studio at up to 1080p
• Input actions logged simultaneously with millisecond-precision timestamps via custom recording software
• All streams temporally synchronized to the same system clock
• Real human gameplay (not scripted or automated)

Production Data Service

This is a demo dataset. We offer large-scale game video collection services:

• Any game title — PC, console (via capture card), mobile
• Hundreds of hours of synchronized gameplay + action data
• Custom annotation layers: game state extraction, object detection, event segmentation
• Multiple players for behavioral diversity
• Monthly capacity: 100,000+ hours

Contact

• Email: simon.su@obaydata.com
• Website: obaydata.com
• Company: New Oriental Bay Limited

Citation

@dataset{obaydata2026worldmodel,
  title={World Model Gameplay Recording},
  author={OBayData Team},
  year={2026},
  url={https://huggingface.co/datasets/obaydata/world-model-gameplay-recording},
  publisher={Hugging Face}
}

License

CC BY-NC 4.0