Datasets:

TencentARC
/

GameHorizon-Data

license: apache-2.0
task_categories:
  - video-text-to-text
language:
  - en
tags:
  - gameplay
  - video
  - instruction-following
  - long-horizon
  - agents
pretty_name: GameHorizon-Data
size_categories:
  - n<1K

GameHorizon-Data

The data home of GameHorizon Suite — a unified data and evaluation suite for gameplay capabilities across temporal horizons. This repository hosts the AAA gameplay recordings, the temporally aligned keyboard-mouse action streams, the multi-horizon instruction annotations, and the benchmark question sets.

📄 Paper: arXiv:2609.25001
💻 Code: TencentARC/GameHorizon
🌐 Project page: https://gamehorizon-suite.github.io

🚀 Release plan

Everything is released here progressively. The annotator demo and the keybind maps are live now; the rest is being prepared.

🎬 GameHorizon-Annotator-Demo/ — six annotated scenes to run the pipeline on
⌨️ keybinds/ — key-to-action maps for all 21 titles
📊 GameHorizon-Bench (offline) — the standardized question sets behind the three primary tasks and the diagnostic variants
📊 GameHorizon-Bench (online) — task specifications and environment states for the stepwise online track
🎯 GameHorizon-Data (full) — 5,000 hours of human gameplay across 21 titles: 4,571 videos at 60 fps, 411.03 M keyboard-mouse events, and 6,184,036 instructions (5,947,588 short-horizon, 189,158 medium-horizon, 47,290 long-horizon)

Watch this repository, or the GitHub repo, to follow progress.

🎬 `GameHorizon-Annotator-Demo/`

Six annotated gameplay scenes, one folder each. Every folder holds the recording, the per-frame keyboard-mouse stream, and the three levels of instructions.

Scene	Game	Frames	Span	Level 1 / 2 / 3
`ApexLegends`	Apex Legends	9,972	603 s	191 / 7 / 3
`Cyberpunk2077`	Cyberpunk 2077	5,253	337 s	94 / 4 / 1
`EldenRing`	Elden Ring	9,884	632 s	419 / 8 / 2
`EldenRing_NightReign`	Elden Ring Nightreign	4,843	311 s	143 / 4 / 2
`GTA5`	GTA V	4,901	313 s	74 / 4 / 1
`Minecraft`	Minecraft	5,237	331 s	92 / 4 / 1

GameHorizon-Annotator-Demo/<Scene>/
├── <recording>.mkv                          # full session, 60 fps
├── <scene>_annotations_per_frame.json       # per-frame keyboard and mouse events
├── level_1_instructions.json                # short-horizon operations
├── level_2_instructions.json                # medium-horizon goals
└── level_3_instructions.json                # long-horizon strategies

*_annotations_per_frame.json — one entry per sampled frame under frames, with the scene's bounds and event counts at the top level.

{
  "scene_name": "...", "interval_start_ns": 0, "interval_end_ns": 0,
  "num_frames": 5237, "stats": {},
  "frames": [{
    "frame_index": 42,
    "timestamp_ns": 1784308897688733400,
    "next_frame_timestamp_ns": 1784308897751233400,
    "width": 1920, "height": 1080,
    "keyboard_events":  [{"timestamp_ns": 0, "event_type": "press", "vk": 32, "key_name": "SPACE"}],
    "raw_mouse_events": [{"timestamp_ns": 0, "dx": -6, "dy": 6,
                          "button_flags": "0", "button_data": 0, "button_flag_names": []}],
    "mouse_events":     [{"timestamp_ns": 0, "event_type": "move", "x": 1279, "y": 822}]
  }]
}

raw_mouse_events is the Windows RawInput stream: dx/dy are relative deltas (dx > 0 rightward, dy > 0 downward) and button_flag_names holds RI_MOUSE_* constants. mouse_events is the derived view of the same input — move with absolute x/y, click with button/pressed, scroll with dx/dy.

Timestamps are absolute wall-clock nanoseconds shared with the recording, which is how frames and actions stay aligned. The annotation covers one scene out of the full session, located through the video's creation_time, so keep the recording intact.

level_{1,2,3}_instructions.json — one entry per clip under items, in temporal order.

{
  "schema_version": 2, "game_id": "minecraft", "level": 1,
  "model": "gemini-3.5-flash", "prompt_version": "detailed-v2-mouse",
  "items": [{
    "scene": "scene_000000",
    "order": 0,
    "level": 1,
    "duration_s": 8.41,
    "instruction": "Strafe right and jump along the crimson cliffside in the Nether ...",
    "source_scenes": [],
    "sampled_frame_refs": ["frames_level_1/scene_000000/000000.jpg"],
    "status": "ok"
  }]
}

source_scenes lists the lower-level clips a Level-2 or Level-3 instruction was merged from, which is what links the three horizons into one pyramid. sampled_frame_refs records the frames the model saw.

⌨️ `keybinds/`

Key-to-action maps for all 21 titles, one JSON per game: a flat mapping from raw KEY_* / RI_MOUSE_* event names to their in-game meaning.

{
  "KEY_W": "Move Forward",
  "LCONTROL": "Sprint",
  "KEY_E": "Open/Close Inventory",
  "RI_MOUSE_LEFT_BUTTON_DOWN": "Attack / Destroy Block"
}

apex_legends · assassins_creed · black_myth_wukong · cyberpunk_2077 · delta_force · eldenring · eldenring_nightreign · escape_from_tarkov · genshin_impact · gta5 · honor_of_kings_world · minecraft · neverness_to_everness · palworld · pubg · red_dead_redemption_2 · roco_kingdom_world · the_witcher_3 · valorant · watch_dogs_2 · wuthering_waves

📥 Download

hf download TencentARC/GameHorizon-Data --repo-type dataset --local-dir ./GameHorizon-Data

# or a subset
hf download TencentARC/GameHorizon-Data --repo-type dataset \
  --include 'keybinds/*' --local-dir ./GameHorizon-Data

📜 License

Apache-2.0. The gameplay recordings are released for research use; all game content remains the property of its respective publishers.

📖 Citation

@article{GameHorizonSuite2026,
      title={GameHorizon Suite: Multi-Horizon Data and Evaluation in Gameplay},
      author={Yiran Wang and Xingyilang Yin and Junfu Pu and Guangzhi Wang and Kaifeng Li and Mingyu Ouyang and Huiqiang Sun and Lingen Li and Cheng Cheng and Wangbo Yu and Honghao Chen and Xiaodong Cun and Chi-Man Pun and Zhiguo Cao and Ying Shan},
      year={2026},
      journal={arXiv preprint arXiv:2609.25001},
      url={https://arxiv.org/abs/2609.25001},
}