The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: TypeError
Message: Couldn't cast array of type
struct<20: struct<informative_references: int64, informative_reward_event: int64, informative_uniform: int64, mean_true_return_range: double, mean_unique_true_returns: double>, 40: struct<informative_references: int64, informative_reward_event: int64, informative_uniform: int64, mean_true_return_range: double, mean_unique_true_returns: double>>
to
{'10': {'informative_references': Value('int64'), 'informative_reward_event': Value('int64'), 'informative_uniform': Value('int64'), 'mean_true_return_range': Value('float64'), 'mean_unique_true_returns': Value('float64')}, '20': {'informative_references': Value('int64'), 'informative_reward_event': Value('int64'), 'informative_uniform': Value('int64'), 'mean_true_return_range': Value('float64'), 'mean_unique_true_returns': Value('float64')}, '5': {'informative_references': Value('int64'), 'informative_reward_event': Value('int64'), 'informative_uniform': Value('int64'), 'mean_true_return_range': Value('float64'), 'mean_unique_true_returns': Value('float64')}}
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2312, in cast_table_to_schema
cast_array_to_feature(
~~~~~~~~~~~~~~~~~~~~~^
table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
feature,
^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1861, in wrapper
return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
~~~~^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2158, in cast_array_to_feature
raise TypeError(f"Couldn't cast array of type\n{_short_str(array.type)}\nto\n{_short_str(feature)}")
TypeError: Couldn't cast array of type
struct<20: struct<informative_references: int64, informative_reward_event: int64, informative_uniform: int64, mean_true_return_range: double, mean_unique_true_returns: double>, 40: struct<informative_references: int64, informative_reward_event: int64, informative_uniform: int64, mean_true_return_range: double, mean_unique_true_returns: double>>
to
{'10': {'informative_references': Value('int64'), 'informative_reward_event': Value('int64'), 'informative_uniform': Value('int64'), 'mean_true_return_range': Value('float64'), 'mean_unique_true_returns': Value('float64')}, '20': {'informative_references': Value('int64'), 'informative_reward_event': Value('int64'), 'informative_uniform': Value('int64'), 'mean_true_return_range': Value('float64'), 'mean_unique_true_returns': Value('float64')}, '5': {'informative_references': Value('int64'), 'informative_reward_event': Value('int64'), 'informative_uniform': Value('int64'), 'mean_true_return_range': Value('float64'), 'mean_unique_true_returns': Value('float64')}}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.
NFWM World-Model Datasets
This repository contains the HDF5 datasets used for NFWM, LeWM, and SMWM training and evaluation across PushT, Reacher, Cube, TwoRooms, LIBERO-90, stochastic Teleport TwoRooms, Atari Boxing, and Atari Breakout.
Repository layout
| Task | Hub path | Local filename | Contents |
|---|---|---|---|
| PushT | datasets/pusht/pusht_expert_train.h5 |
pusht_expert_train.h5 |
2,336,736 transitions from 18,685 episodes, with 224 x 224 RGB observations |
| Reacher | datasets/reacher/reacher.h5 |
reacher.h5 |
2,010,000 transitions from 10,000 episodes, with 224 x 224 RGB observations |
| Cube | datasets/cube/cube_single_expert.h5 |
cube_single_expert.h5 |
2,010,000 transitions from 10,000 episodes, with 224 x 224 RGB observations |
| TwoRooms | datasets/tworooms/tworoom.h5 |
tworoom.h5 |
920,809 transitions from 10,000 episodes, with 224 x 224 RGB observations |
| LIBERO-90 | datasets/libero/libero_90_all_7d.h5 |
libero_90_all_7d.h5 |
669,043 steps from 4,500 demonstrations across 90 tasks, with 128 x 128 RGB observations and 7D actions |
| Teleport TwoRooms V5 train | datasets/teleport/tworoom_teleport_v5_exactgoal_train.h5 |
tworoom_teleport_v5_exactgoal_train.h5 |
711,327 steps from 20,000 episodes, with 224 x 224 RGB observations and 2D actions |
| Teleport TwoRooms V5 eval | datasets/teleport/tworoom_teleport_v5_exactgoal_eval.h5 |
tworoom_teleport_v5_exactgoal_eval.h5 |
1,997 steps from 50 held-out episodes arranged as 25 stochastic pairs |
| Atari Boxing train | datasets/boxing/boxing_rgb96_seed1_train.h5 |
boxing_rgb96_seed1_train.h5 |
904,014 transitions from 751 episodes |
| Atari Boxing test | datasets/boxing/boxing_rgb96_seed1_test.h5 |
boxing_rgb96_seed1_test.h5 |
101,287 transitions from 84 episodes |
| Atari Breakout train | datasets/breakout/breakout_rgb96_seed1_train.h5 |
breakout_rgb96_seed1_train.h5 |
921,755 transitions from 1,121 episodes |
| Atari Breakout test | datasets/breakout/breakout_rgb96_seed1_test.h5 |
breakout_rgb96_seed1_test.h5 |
98,219 transitions from 125 episodes |
The LIBERO-90 file contains 50 demonstrations per task. Actions use
[dx, dy, dz, d_axis_angle_x, d_axis_angle_y, d_axis_angle_z, gripper] and no
language conditioning is provided to the world model.
Teleport TwoRooms V5 uses frame skip 5, history size 3, speed-5 expert motion, and an exact-goal training set. The 20,000 training episodes combine 9,000 paired successful episodes, 9,000 counterfactual episodes, and 2,000 collision episodes. The evaluation split contains 25 paired stochastic outcomes and uses canonical exact target rendering. Hub evaluation files are physically materialized and contain no machine-local HDF5 virtual-dataset references.
Boxing observations are 96 x 96 RGB frames from ALE/Boxing-v5 with frame
skip 5 and sticky actions disabled. The split seed is 20260819, and the test
fraction is 0.1. Supporting split manifests and validation reports are stored
next to the train and test files.
The datasets/boxing/evaluation/ directory contains the counterfactual
action-ranking development sets and the scene-coverage image set used by the
Boxing evaluation and decoder workflows.
Breakout observations are 96 x 96 RGB frames from ALE/Breakout-v5 with frame
skip 5 and sticky actions disabled. The split seed is 20260819, and the test
fraction is 0.1. Supporting split manifests and validation reports are stored
next to the train and test files.
HDF5 schemas
The robotics and control datasets use flat arrays such as pixels, action,
episode offsets, and task-specific state fields. LIBERO additionally stores
task IDs, task names, per-episode goal indices, and episode offsets. Teleport
includes agent and target positions, proprioception, rewards, termination
flags, and environment state fields. The Atari train/test files use one group
per episode. Each episode stores observations[T+1] and aligned transition
arrays for actions, rewards, termination flags, truncation flags, and
timesteps.
Integrity
Task-specific SHA256SUMS files are generated from the exact uploaded files.
Verify downloaded files from the corresponding task directory with:
sha256sum -c SHA256SUMS --ignore-missing
Provenance
The PushT, Cube, Reacher, and TwoRooms files were obtained from the public LeWM dataset collection. The local files are uploaded here in their decompressed HDF5 form so that the same paths can be used directly on another cluster.
The LIBERO-90 file was converted from the LIBERO demonstration suite into the flat Stable World Model format used by NFWM. The Teleport TwoRooms V5 files were generated from the stochastic compact TwoRooms environment for controlled multimodal world-model experiments.
The Boxing files were prepared for the NFWM/LeWM/SMWM Boxing experiments from Atari Learning Environment trajectories. The action-ranking files are development benchmarks derived from the held-out Boxing split.
The Breakout files were prepared for the NFWM/LeWM/SMWM Breakout experiments from Atari Learning Environment trajectories. The train and test HDF5 files are materialized episode-level splits of the collected trajectory dataset.
Users are responsible for checking and complying with the licenses and terms of the upstream environments, datasets, and Atari assets before redistribution or commercial use.
Checkpoints
Dataset files are uploaded first. Checkpoint publication is intentionally
tracked separately under checkpoints/ and may later move to a dedicated
Hugging Face model repository.
- Downloads last month
- -