The dataset viewer is not available for this subset.
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 84, in _split_generators
raise ValueError(
...<2 lines>...
)
ValueError: The TAR archives of the dataset should be in WebDataset format, but the files in the archive don't share the same prefix or the same types.
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.
ShowNow Multi-Future Dataset
Matched multi-future video data for action-conditioned, interactive world models, released with ShowNow: Fine-Grained Speculative Streaming for Responsive World Models.
Each sample is a fork: a short first-person history followed by eight futures. All eight futures start from the same restored game state and differ only in the action taken. The futures are matched counterfactuals: any difference between them is caused by the action, not by a different starting point.
| subset | forks | futures per fork | action schedule in the future | use in the paper |
|---|---|---|---|---|
held |
10,241 | 8 | one key held for the whole 3 s future | training |
switch |
11,471 | 8 | each branch starts with its own key, then switches among the 8 keys every 0.25-1.5 s (25 % held throughout) | training |
eval |
252 | 8 | as switch, new game states (different generator seed) |
validation / test (eval_split.json) |
Actions (one per branch): W forward, A strafe left, S backward, D strafe right, TURN_L, TURN_R, ATTACK, JUMP.
Format
WebDataset-style tar shards (<subset>/<subset>-NNNNN.tar, 1000 forks per shard). Each fork is a folder <fork_id>/
containing:
history.mp4: the history before the fork.future_<ACTION>.mp4x 8: the 3 s future (96 frames) for each action.meta.json, which holds:- engine and map: ViZDoom 1.3.1 / GZDoom,
freedoom2.wad, map, skill, seed; - camera: intrinsics, plus a camera-to-world pose per frame (OpenCV convention, metres);
- per frame: the button vector and the game state;
- the fork state (position, yaw, velocity, health, weapon, ammo), and nearby and visible monsters;
- per-future captions and effect summaries;
- consistency checks of the state restoration.
- engine and map: ViZDoom 1.3.1 / GZDoom,
preview.jpg, where present.
Video: 1280x720, 32 fps (frames sampled at the nearest engine tic of t = n/32 s; the engine runs at 35 tics/s), horizontal FOV 130 degrees. The paper trains at 16 fps and 864x480 (temporal and spatial resampling of these files).
manifest.json lists the forks and the SHA-256 of every shard. eval_split.json gives the evaluation splits:
e2_val (48 forks), test_e2 (96), and one-minute rollout scenes scenes (10) / test_e3 (20).
How the data were generated
An exploring agent plays Freedoom2 maps MAP01-MAP28 in ViZDoom with god mode. At each sampled fork point the game state
is saved. Each of the eight action branches is then rendered from the restored state. Restoration is verified: each
branch must reproduce the saved snapshot; forks that fail are discarded. Forks with damage in the history are excluded,
and so are forks facing a wall or with a too-dark history (thresholds are in meta.json under checks). The held and
switch subsets were used, mixed, for all training stages in the paper. eval uses training maps with a different
generator seed and was used only for evaluation.
Loading
from huggingface_hub import hf_hub_download
import tarfile
p = hf_hub_download("songyiren/ShowNow-MultiFuture", "held/held-00000.tar", repo_type="dataset")
tarfile.open(p).extractall("shownow_mf") # -> shownow_mf/<fork_id>/{history.mp4, future_*.mp4, meta.json}
License
The videos are rendered with ViZDoom from the Freedoom assets. Freedoom is distributed under a BSD-style license, and this dataset is released under BSD-3-Clause.
Citation
@article{shownow2026,
title = {ShowNow: Fine-Grained Speculative Streaming for Responsive World Models},
author = {Song, Yiren and Fang, Guian and Shou, Mike Zheng},
year = {2026}
}
- Downloads last month
- 1