Datasets:
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 80, 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.
UPBench-Error (filtered)
A high-confidence subset of LCZZZZ/UPBench-Error's
gt_error split, for testing whether an assistant watching a video live — holding only one short standing
request from the user — notices a mistake as it happens and says what is wrong.
250 help points across 223 videos, filtered from 7826 help points across 3179 videos (3.2% kept).
| Modality | Videos | Help points |
|---|---|---|
non-GUI (task_family: G) |
120 | 145 |
GUI (task_family: T) |
103 | 105 |
| Criterion type | Help points |
|---|---|
| error_correction | 194 |
| correction_confirm | 39 |
| next_step_guidance | 12 |
| risk_alert | 5 |
Sources: holoassist 101, guide 66, a11y_cua 18, captaincook4d 14, mlfoundations 8, clawbench_v2 / cua_debugger / look_before_you_click 3 each, ego_mc_bench / impact / pango 2 each, egooops 1.
Videos are the parent dataset's originals, byte for byte — same resolutions (640×360 up to 3840×2160, depending on source) and same 30 fps. Nothing was re-encoded.
Why filter
Every annotation in the parent dataset was produced by gemini-3.6-flash three-vote convergence with no
human verification. Sampling frames against the labels shows a high rate of items that cannot be answered
from the video at all.
How it was filtered
- Decidability judge — given only the video and the user's standing request, could one tell that this is a mistake? Dropped when recognising it requires a norm the assistant cannot have (a recipe quantity or duration, a manual's step order, which numbered bin a part belongs in), when the deciding detail is off-frame or unresolvable, when nothing is actually wrong, when it is plain accident detection, or when the named action does not occur in the video. GUI items were judged on their own rubric, adding contrived / not worth interrupting, legitimate design choice, and recording carries no information.
- Two-pass consensus — the judge was run twice independently, pinned to a single model; only points kept by both passes survive. (Test-retest Jaccard on KEEP was 36% before pinning the model and 63% after, so this step matters.)
- Blind-guess probe — with no video, only the task and the standing request, a model predicts what error is likely; items whose reference answer a blind guess would match are dropped as answerable without watching.
- Semantic dedup — embedding clustering (complete linkage, cosine distance 0.12), one item per cluster.
- Timing relocation — original timestamps were not trusted. Each window was re-derived from the video: when the mistake first becomes visible, and when saying it stops being useful. Where the person corrects themselves, the window ends before that.
Windows in window_unified are the relocated ones (at_frame and gt_error.quality.timing.optimal_s
follow). Every other field is carried over verbatim from the parent dataset — the schema is identical,
field for field, 85 fields, nothing added and nothing dropped.
Layout
gt_error/
├── metadata/manifest-000.jsonl all 223 records
├── metadata/G-000.jsonl non-GUI subset
├── metadata/T-000.jsonl GUI subset
├── summary.json
└── videos/videos_00..04.tar 223 mp4 files, original encodings
Unpack videos with for f in gt_error/videos/*.tar; do tar -xf "$f" -C gt_error/videos/; done.
Caveats
- holoassist dominates (101 of 223 videos). It is 77% of the parent dataset's help points, and its device-operation scenarios (cameras, coffee machines, circuit breakers) satisfy the "the evidence is visible on screen" bar more often than furniture assembly does.
- ego_mc_bench and impact almost vanish (2 videos each). Their annotations lean on recipe quantities and on a disassembly protocol's bin numbering — decidable only from documents the assistant cannot have. This is a property of those annotation styles, not a filtering artifact.
- Filtering was done by VLM judges calibrated against a human-annotated sample, not by exhaustive human review. Each surviving item records why it was kept.
- Downloads last month
- 62