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 89, in _split_generators
pa.Table.from_pylist(cast_to_python_objects([example], only_1d_for_numpy=True))
~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "pyarrow/table.pxi", line 2044, in pyarrow.lib._Tabular.from_pylist
File "pyarrow/table.pxi", line 6401, in pyarrow.lib._from_pylist
return cls.from_arrays(arrays, names, metadata=metadata)
File "pyarrow/table.pxi", line 4868, in pyarrow.lib.Table.from_arrays
converted_arrays = _sanitize_arrays(arrays, names, schema, metadata,
File "pyarrow/table.pxi", line 1611, in pyarrow.lib._sanitize_arrays
converted_arrays = _schema_from_arrays(arrays, names, metadata,
File "pyarrow/table.pxi", line 1592, in pyarrow.lib._schema_from_arrays
val = array(val)
File "pyarrow/array.pxi", line 400, in pyarrow.lib.array
File "pyarrow/array.pxi", line 46, in pyarrow.lib._sequence_to_array
chunked = GetResultValue(
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
return check_status(status)
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowInvalid: cannot mix list and non-list, non-null values
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.
UniNavBench
UniNavBench contains long-horizon embodied navigation and Active Embodied Question Answering (Active-EQA) episodes in dynamic ProcTHOR indoor scenes. Scheduled activities relocate objects and change their states over 1–3 simulated days. Each episode contains 10 sub-tasks in the same scene.
This repository contains evaluation annotations, dynamic schedules and goal images. Run them with the accompanying UniNavBench Python evaluation package. ProcTHOR houses and the AI2-THOR simulator are obtained separately by the evaluation package; their upstream licenses apply to those resources.
Splits
| Download split | Scenes | Navigation episodes | Active-EQA episodes |
|---|---|---|---|
| train | 500 | 3 × 1,000,000 | 3 × 1,000,000 |
| val | 200 | 3 × 400,000 | 3 × 400,000 |
| test | 100 | 3 × 200,000 | 3 × 200,000 |
| mini_test | 3 | 3 × 6,000 | 3 × 6,000 |
Each split has three dynamic levels: static (no scheduled changes), small (smaller objects and states change), and all (furniture can move as well).
mini_test contains test scenes 0, 4 and 6, including all their episodes, levels, schedules and required goal images. These scenes remain in the full test split. It is a convenience subset for installation checks and agent development, not an independent held-out split. Results on mini_test should not be presented as full-test benchmark scores.
Download
After installing the evaluation package from the UniNavBench source repository:
uninavbench download --split mini_test
uninavbench download --split train val test
The default extraction directory is ~/.cache/uninavbench/data; override it with --data_pth DIR and use UNINAVBENCH_DATA=DIR for evaluation. Every tar shard is checked against the SHA-256 in manifest.json.
The Python package has not yet been published to PyPI. Use an up-to-date source checkout that supports the mini_test download option.
To download without installing the evaluation package:
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="yhncao/UniNavBench",
repo_type="dataset",
allow_patterns=["manifest.json", "README.md", "mini_test-*.tar"],
local_dir="UniNavBench_release",
)
Then use the evaluation package to verify and extract the local shards:
uninavbench download --split mini_test --from_dir UniNavBench_release
Run a small check
mini_test archives preserve the original test/ paths and ProcTHOR test scene IDs. Evaluate with -p test, not -p mini_test:
uninavbench run -p test -i 0 -l all -t nav -es 0 -ee 1
The default agent moves randomly. Pass -a /path/to/your_agent.py to evaluate your own agent. Active-EQA additionally requires a configured judge; exact matching is available for offline checks.
Archive format
The dataset uses tar shards and a manifest rather than tabular Parquet files. Download the archives and use the evaluator; automatic table previews are not provided.
train/<scene>/episodes/<level>/{nav,eqa}.json.gz
train/<scene>/dynamic_strategy/strategy_final_{small,all}.json
attributes/train/<scene>/images/<goal-image>.png
val and test follow the same layout. The mini_test-000.tar shard contains test/ and attributes/test/ paths. Its manifest entry uses split: mini_test and source_split: test. The full splits use 25 scenes per shard; mini_test contains all 3 scenes in one shard. Scene IDs refer to the original ProcTHOR split and are not renumbered.
Release status
This is an initial research release. Evaluation checks cover a small sample, not all scenes. A known scene-0 small-level replay check found a roughly 6.4 cm vertical discrepancy for a garbage can, above the current 5 cm positional tolerance; the sampled oracle tasks still succeeded. Scene replay consistency and environment version pinning remain under investigation. Use the evaluation code and report its version and simulator environment alongside results.
Completeness repair (2026-11-14)
All train, val, test and mini_test shards are included in manifest version 2026.11.14.1. The six train/220 and train/229 navigation files each contain 2000 episodes of 10 subtasks. A generator shortfall was repaired by resampling 467 existing same-scene, same-level relation navigation tasks with full-day visibility evidence and conservative schedule checks. Original task prefixes and episode metadata are preserved; the 43 added tail episodes contain relation navigation only, restoring 4000 subtasks of each of the five navigation types per file. Per-task provenance is provided in the repair report. These are replacement samples, not a recovery of original random selections.
- Downloads last month
- 58