Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 88, in _split_generators
inferred_arrow_schema = pa.concat_tables(pa_tables, promote_options="default").schema
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "pyarrow/table.pxi", line 6268, in pyarrow.lib.concat_tables
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowTypeError: Unable to merge: Field json has incompatible types: struct<frame_id: int64, simulation_step: int64, engine_frame_counter: int64, time_seconds: double, delta_seconds: double, camera_mode: string, width: int64, height: int64, semantic_encoding: string, semantic_background_min_depth_m: int64, background_stencil_changes: int64, label_invalid_pixels: int64, orthographic: bool, fov_x_deg: double, ortho_width_cm: double, camera_world_cm: list<item: double>, camera_pitch_yaw_roll: list<item: double>, world_to_view: list<item: double>, projection: list<item: double>, world_to_clip: list<item: double>, actors: list<item: struct<role: string, class_id: int64, action: string, health: int64, position_cm: list<item: double>, velocity_cm_s: list<item: int64>, camera_hidden: bool, bones: list<item: null>>>, whitebox: struct<representation: string, simulation_step: int64, world_to_clip: list<item: double>, capsules: list<item: struct<role: string, from_bone: string, to_bone: string, head_extension: bool, radius_cm: double, a_cm: list<item: double>, b_cm: list<item: double>, camera_hidden: bool>>>, frame_fps: int64, simulation_hz: int64, episode: int64, camera_variant: int64, mirrored_world: bool, personality: string, appearance: string, events: list<item: null>, characters: list<item: struct<role: string, model: string, material_variant: string, scale: double, height_cm: double, capsule_radius_cm: double, capsule_half_height_cm: double, skin_hue_sat_bright: list<item: double>>>, v9: struct<look: int64, look_name: string, capsule_labels: bool, frame_offset_cm: list<item: int64>, characters: list<item: struct<role: string, class_id: int64, capsules: list<item: struct<name: string, a: list<item: double>, b: list<item: double>, r: double>>>>>, action: struct<w: int64, a: int64, s: int64, d: int64, attack: int64, reset: bool, block: int64, block_start: bool, bot_mode: string, player_stamina: int64>> vs struct<frame_id: int64, simulation_step: int64, engine_frame_counter: int64, time_seconds: double, delta_seconds: double, camera_mode: string, width: int64, height: int64, semantic_encoding: string, semantic_background_min_depth_m: int64, background_stencil_changes: int64, label_invalid_pixels: int64, orthographic: bool, fov_x_deg: double, ortho_width_cm: double, camera_world_cm: list<item: double>, camera_pitch_yaw_roll: list<item: double>, world_to_view: list<item: double>, projection: list<item: double>, world_to_clip: list<item: double>, actors: list<item: struct<role: string, class_id: int64, action: string, health: int64, position_cm: list<item: double>, velocity_cm_s: list<item: double>, camera_hidden: bool, bones: list<item: null>>>, whitebox: struct<representation: string, simulation_step: int64, world_to_clip: list<item: double>, capsules: list<item: struct<role: string, from_bone: string, to_bone: string, head_extension: bool, radius_cm: double, a_cm: list<item: double>, b_cm: list<item: double>, camera_hidden: bool>>>, frame_fps: int64, simulation_hz: int64, episode: int64, camera_variant: int64, mirrored_world: bool, personality: string, appearance: string, events: list<item: null>, characters: list<item: struct<role: string, model: string, material_variant: string, scale: double, height_cm: double, capsule_radius_cm: double, capsule_half_height_cm: double, skin_hue_sat_bright: list<item: double>>>, v9: struct<look: int64, look_name: string, capsule_labels: bool, frame_offset_cm: list<item: int64>, characters: list<item: struct<role: string, class_id: int64, capsules: list<item: struct<name: string, a: list<item: double>, b: list<item: double>, r: double>>>>>, action: struct<w: int64, a: int64, s: int64, d: int64, attack: int64, reset: bool, block: int64, block_start: bool, bot_mode: string, player_stamina: int64>>: Unable to merge: Field actors has incompatible types: list<item: struct<role: string, class_id: int64, action: string, health: int64, position_cm: list<item: double>, velocity_cm_s: list<item: int64>, camera_hidden: bool, bones: list<item: null>>> vs list<item: struct<role: string, class_id: int64, action: string, health: int64, position_cm: list<item: double>, velocity_cm_s: list<item: double>, camera_hidden: bool, bones: list<item: null>>>: Unable to merge: Field item has incompatible types: struct<role: string, class_id: int64, action: string, health: int64, position_cm: list<item: double>, velocity_cm_s: list<item: int64>, camera_hidden: bool, bones: list<item: null>> vs struct<role: string, class_id: int64, action: string, health: int64, position_cm: list<item: double>, velocity_cm_s: list<item: double>, camera_hidden: bool, bones: list<item: null>>: Unable to merge: Field velocity_cm_s has incompatible types: list<item: int64> vs list<item: double>: Unable to merge: Field item has incompatible types: int64 vs double
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
GameCrafter v10 Paired Data (capsule characters)
Paired training data for an action-conditioned 2.5D game world model whose control signal is a per-pixel numeric semantic map. It follows GameCrafter-Data-v8 and contains every production run captured since the player and the boss are drawn as articulated capsule whiteboxes in the labels:
| Subset | What it is | Runs | Frames | Game time | 81-frame clips |
|---|---|---|---|---|---|
v10_real |
photoreal outdoor biomes, telephoto camera ("realistic cinematic") | 116 | 556,800 | 9.67 h | 5,800 |
v9_stylized |
the 10 game scenes as v8 cross-scene worlds, with a random rendering style per camera variant | 140 | 672,000 | 11.67 h | 7,000 |
| total | 256 | 1,228,800 | 21.33 h | 12,800 |
Each frame stores the rendered RGB, a pixel-aligned semantic id map, a structural whitebox render, a ready-made condition image (semantic palette + code-drawn capsule characters), the player's 5-dim action (W, A, S, D, attack) and per-frame camera / actor / capsule metadata. Size: 482 GB.
What changed since v8
- Capsule characters in the labels (v9). In
ids.pngthe player (class 10) and the boss (class 11), with their weapons, are drawn as the articulated capsule proxies of the whitebox pass, not as mesh silhouettes. Every frame's json also stores the capsule segments (world cm,v9.characters), so the character condition can be redrawn by code without Unreal.cond.pngis that condition, written for every frame: the semantic palette with the player / boss capsules shaded per role. - Clips on a cut grid (v9). Looks, cameras and episodes only change between frame
96k+80and the next, and every run has exactly 50 clips of 81 frames starting at multiples of 96 (clips_81.json). No clip contains a style, camera or episode change. The 15 frames between two clips are not part of any clip (the first frame after a switch can still show unstreamed textures). - Closer camera. The frame is 900–1600 cm wide at the player (12% of views 2000–3000 cm). In
v10_realthe player always stays in the centre of the image and the boss may leave the frame; inv9_stylizedthe camera eases towards the player–boss midpoint during fights, andclips_81.jsonmarks the clips whose camera never left the player (centered). v9_stylized: a new look per camera variant. On top of the v8 cross-scene worlds, every camera variant (6–12 per run, always on the cut grid) gets a random RGB look (labels and whitebox unchanged): post-process style, material randomisation, render tier, texture mip bias, colour grade and outline.v10_real: photoreal biomes. The authored set dressing is hidden and every look (one per clip, every 6 s of game time) re-scatters CC0 photo-scanned assets (Poly Haven) on the scene's walkable ground in one of 12 biome variants, with matching photo-scanned ground / slope materials, foliage season tint, snow / moss / dust / sand cover on the models, wetness and a light preset (including night and storm). The camera is a perspective telephoto (FOV 22–40°) with the same framing as the 2.5D camera.
At a glance
v10_real |
v9_stylized |
|
|---|---|---|
| Runs | 116 × 5 min (4,800 frames) | 140 × 5 min (4,800 frames) |
| Frames / game time | 556,800 / 9.67 h | 672,000 / 11.67 h |
| 81-frame clips (frames inside clips) | 5,800 (8.16 h) | 7,000 (9.84 h) |
| Camera | perspective telephoto, FOV 22–40° | isometric orthographic |
| Looks | a new biome / light / camera every 6 s (one per clip) | 6–12 per run, one per camera variant |
| Dynamic events | none | 6,096 (spawn / replace / delete) |
| Clips with the player centred on every frame | 5,800 | 4,305 |
| Episodes | 925 | 1,124 |
| Label check | 0 invalid semantic pixels | 0 invalid semantic pixels |
| Resolution | 1280 × 704 | 1280 × 704 |
Previews: previews/v10_real_preview.mp4 and previews/v9_stylized_preview.mp4 show two clips of one run per scene (RGB | condition).
Per-frame contents
Each frame is one WebDataset sample inside a tar shard; the key is the 6-digit frame index within its run.
| File in tar | Content |
|---|---|
{frame}.jpg |
RGB, 1280×704, JPEG (rendered at 2× and downsampled) |
{frame}.ids.png |
Semantic id map, single-channel uint8, lossless, pixel-aligned with the RGB; player / boss as capsules |
{frame}.whitebox.jpg |
Whitebox render: structural scene geometry plus the articulated character capsules |
{frame}.cond.png |
Condition image: semantic palette for the scene classes + player (blue) / boss (red) capsules shaded from the capsule segments |
{frame}.json |
Per-frame metadata: camera pose, world_to_view, projection, world_to_clip, fov_x_deg / ortho_width_cm, actors (role, state, health, position, velocity, bones with pixel positions), action, episode, camera_variant, v9 (look, capsules) |
Semantic classes
| id | class | id | class | id | class |
|---|---|---|---|---|---|
| 0 | sky / background | 3 | architecture | 6 | prop |
| 1 | terrain (incl. grass, debris) | 4 | rock | 10 | player |
| 2 | paving (roads, floors) | 5 | vegetation (plants, trees, logs) | 11 | boss |
Pixel share (6 frames per run):
| Class | v9_stylized |
v10_real |
|---|---|---|
| terrain | 18.9% | 92.3% |
| paving | 37.0% | 0.0% |
| architecture | 22.2% | 0.0% |
| vegetation | 14.1% | 3.8% |
| rock | 3.1% | 1.8% |
| prop | 3.0% | 0.0% |
| boss | 0.9% | 1.1% |
| player | 0.7% | 0.9% |
Index files per run
data/<subset>/<scene>/<run>/frames-000000.tar frames 0–1599 (WebDataset shard)
data/<subset>/<scene>/<run>/frames-001600.tar frames 1600–3199
data/<subset>/<scene>/<run>/frames-003200.tar frames 3200–4799
data/<subset>/<scene>/<run>/index/
actions.npy (N,5) uint8 [w,a,s,d,attack]; row i = keys held from frame i to i+1
actions_block_0.25s.npy (N/4,5) one action prompt per 0.25 s (keys never change inside a block)
episodes.npy (N,) episode id
actions.csv per-frame actions, bot mode, player / boss state and HP, character model and scale
frames.jsonl compact per-frame metadata
clips_81.json the 50 cut-grid clips: start, end (exclusive), episode, camera_variant, look, look_name, events
clips_81_v9.json the capture's own clip list (end inclusive)
events.json dynamic events of the run (v9_stylized only)
meta.json run statistics (looks, behaviour, characters, v10 biome variants)
capture_result.json capture settings and the per-look log
scene_actors.json static-mesh inventory of the host scene
summary_v10.csv / summary_v10.json every run and the totals of both subsets
previews/ thumbs/ code/ preview reels, one thumbnail per run (RGB | semantic | condition), capture and post code
Loading a clip
import io, json, tarfile
import numpy as np
from PIL import Image
from huggingface_hub import hf_hub_download
REPO = "teawhite/GameCrafter-Data-v10"
subset, scene, run = "v10_real", "05_Japanese_Spring_Sanctuary", "R10_s5501"
get = lambda p: hf_hub_download(REPO, f"data/{subset}/{scene}/{run}/{p}", repo_type="dataset")
clips = json.load(open(get("index/clips_81.json")))
blocks = np.load(get("index/actions_block_0.25s.npy"))
c = clips[10]; s, e = c["start"], c["end"] # 81 frames, one look (c["look_name"])
def frames(lo, hi): # frames [lo, hi) from the 1,600-frame shards
out = {}
for shard in range(lo // 1600 * 1600, hi, 1600):
with tarfile.open(get(f"frames-{shard:06d}.tar")) as tf:
for m in tf:
key, ext = m.name.split(".", 1)
if lo <= int(key) < hi: out.setdefault(int(key), {})[ext] = tf.extractfile(m).read()
return [out[i] for i in range(lo, hi)]
clip = frames(s, e)
rgb = np.stack([np.asarray(Image.open(io.BytesIO(f["jpg"]))) for f in clip]) # (81,704,1280,3)
cond = np.stack([np.asarray(Image.open(io.BytesIO(f["cond.png"]))) for f in clip]) # condition images
ids = np.stack([np.asarray(Image.open(io.BytesIO(f["ids.png"]))) for f in clip]) # (81,704,1280) uint8 class ids
act = blocks[s // 4 : s // 4 + 20] # (20,5) one prompt per 0.25 s
v10_real: biomes and light
Every clip has its own biome variant and light preset (look_name = photo|<light>|<biome>|pbr|no).
| Biome variant | Clips | Share |
|---|---|---|
rain_wetland |
500 | 8.6% |
dry_savanna |
495 | 8.5% |
rocky_coast |
495 | 8.5% |
sand_beach |
487 | 8.4% |
snow_alpine |
486 | 8.4% |
spring_meadow |
485 | 8.4% |
semi_desert |
480 | 8.3% |
red_canyon |
479 | 8.3% |
conifer_forest |
474 | 8.2% |
rock_badlands |
474 | 8.2% |
autumn_wood |
473 | 8.2% |
mossy_oldwood |
472 | 8.1% |
| Light preset | Clips | Share |
|---|---|---|
overcast |
997 | 17.2% |
golden |
989 | 17.1% |
dusk |
840 | 14.5% |
backlit |
821 | 14.2% |
noon |
777 | 13.4% |
mist |
761 | 13.1% |
storm |
396 | 6.8% |
night |
219 | 3.8% |
v9_stylized: looks
look_name = <post style>|<materials>|<render tier>|<mip bias>|<outline>.
| Post-process style | Clips | Share |
|---|---|---|
none |
2,565 | 36.6% |
sketch |
578 | 8.3% |
pixel |
564 | 8.1% |
comic |
520 | 7.4% |
toon |
508 | 7.3% |
duotone |
487 | 7.0% |
soft |
477 | 6.8% |
paint |
466 | 6.7% |
dither |
419 | 6.0% |
hue |
416 | 5.9% |
| Materials | Clips | Share |
|---|---|---|
orig |
3,331 | 47.6% |
tex |
1,261 | 18.0% |
flat |
876 | 12.5% |
mixed |
838 | 12.0% |
shuffle |
694 | 9.9% |
| Render tier | Clips | Share |
|---|---|---|
full |
4,239 | 60.6% |
nogi |
1,312 | 18.7% |
unlit |
737 | 10.5% |
noshadow |
712 | 10.2% |
Behaviour
Share of frames by bot mode:
| Bot mode | v9_stylized |
v10_real |
|---|---|---|
| explore | 36.3% | 43.2% |
| fight | 15.9% | 18.4% |
| seek_boss | 13.1% | 9.1% |
| dead | 9.1% | 10.0% |
| pause | 5.3% | 5.5% |
| unstick | 6.3% | 1.6% |
| detour | 3.8% | 4.0% |
| retreat | 3.3% | 4.0% |
| backtrack | 2.8% | 2.7% |
| wallhug | 2.0% | 0.2% |
| disengage | 0.9% | 1.3% |
| push_obstacle | 1.0% | 0.1% |
Frames with no key held: 22.5% (v9_stylized), 24.2% (v10_real).
| Scene | v9 runs |
v9 clips |
v10_real runs |
v10_real clips |
v10_real fight share* |
|---|---|---|---|---|---|
| 01 Sunlit Mediterranean Courtyard | 15 | 750 | 12 | 600 | 17.1% |
| 02 Autumn Woodland Shrine | 15 | 750 | 12 | 600 | 17.5% |
| 03 Desert Caravanserai | 14 | 700 | 12 | 600 | 42.9% |
| 04 Snowbound Mountain Outpost | 14 | 700 | 12 | 600 | 19.8% |
| 05 Japanese Spring Sanctuary | 14 | 700 | 12 | 600 | 16.2% |
| 06 Overgrown Tropical Water Temple | 14 | 700 | 12 | 600 | 21.9% |
| 07 Steampunk Foundry District | 14 | 700 | 11 | 550 | 17.4% |
| 08 Gothic Cathedral Quarter | 14 | 700 | 11 | 550 | 13.6% |
| 09 Dense Cyberpunk Neon District | 14 | 700 | 11 | 550 | 28.5% |
| 10 Underwater Research Habitat | 12 | 600 | 11 | 550 | 28.1% |
* Share of frames in which the bot is fighting or retreating.
Capture details
- Unreal Engine 5.8, the 10 GameCrafter scenes. Offline fixed-timestep simulation at 48 Hz, every 3rd step saved: 16 fps of game time independent of wall-clock speed.
- The bot plays through the player's own input path (WASD + attack), explores, seeks and fights the boss, dies and respawns (a new episode). Decisions are made every 4 frames.
v9_stylizedwas captured on 2026-10-20,v10_realon 2026-10-22 (after a fix of the moss cover material; two earlier runs with a flat moss colour are not included).
Known limitations
- Same 10 host scenes and layouts.
v10_realreplaces the set dressing and materials but keeps each scene's walkable ground and routes. v10_realhas no dynamic events, and its semantic maps are dominated by terrain.- RGB-only stylisation in
v9_stylized. Styles, outlines and material swaps exist only in the RGB; the labels, whitebox and condition show the plain scene. - Frames between clips. The 15 frames after each clip may contain a look switch; use
clips_81.jsoninstead of slicing runs freely.
- Downloads last month
- 536