Datasets:

Cognyzer
/

osworld-ubuntu

Error code:   StreamingRowsError
Exception:    TypeError
Message:      Couldn't cast array of type
struct<command: list<item: string>, files: list<item: struct<url: string, path: string>>>
to
{'command': List(Value('string')), 'seconds': Value('int64'), 'window_name': Value('string')}
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 2118, in cast_array_to_feature
                  casted_array_values = _c(array.values, feature.feature)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1863, in wrapper
                  return func(array, *args, **kwargs)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2068, in cast_array_to_feature
                  _c(array.field(name) if name in array_fields else null_array, subfeature)
                  ~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1863, in wrapper
                  return func(array, *args, **kwargs)
                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<command: list<item: string>, files: list<item: struct<url: string, path: string>>>
              to
              {'command': List(Value('string')), 'seconds': Value('int64'), 'window_name': Value('string')}

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.

OSWorld Ubuntu Tasks

Eight OSWorld-format desktop tasks for Ubuntu, each with a deterministic, rule-based evaluator. Built by Cognyzer.

Every task targets a real GUI application rather than a terminal workflow, and every one is graded by inspecting the artefact the agent produced (a WAV file, a spreadsheet, a mail profile, an image) instead of by matching keystrokes.

Tasks

Folder	Application	What the agent has to do	Human steps
`audacity`	Audacity	Trim a track to a 75 second cut, apply fades, prepend silence, mix to mono, export 16 bit PCM WAV	49
`gimp`	GIMP	Scale to 400x400, convert to greyscale then back to RGB, add a 30 px `#2E8B57` border, export a new PNG	49
`libreoffice-calc`	LibreOffice Calc	Insert a row, add a cell comment, two-key sort, derive a Budget Status column, set column widths	45
`libreoffice-impress`	LibreOffice Impress	Replace two template tokens across a ten slide deck without disturbing any other content or creating stray files	40
`libreoffice-writer`	LibreOffice Writer	Password protect a named section, add a comment on a heading, delete one specific existing comment	26
`pinta`	Pinta	Resize a screenshot to 1280x720 and redact a footer strip with an opaque black rectangle	26
`shotcut`	Shotcut	Set twelve interlinked preferences across the main and timeline settings panes	35
`thunderbird`	Thunderbird	Create a folder, build a three condition message filter, and change four notification preferences	40

Layout

tasks/<name>/task.json   task definition
assets/                  files the task configs and evaluators download

A human demonstration trajectory exists for every task, recorded as one pyautogui action per line alongside a screenshot per step. Those bundles are delivered separately and are not part of this repository; the trajectory field in each task.json is the path they unpack to, relative to the task folder.

Task schema

Every task.json uses the same key order:

Field	Notes
`id`	UUID
`instruction`	Natural language request given to the agent
`snapshot`	VM image the task expects
`trajectory`	Where the human trajectory sits, relative to the task folder
`config`	Environment setup run before the agent starts
`evaluator`	`postconfig`, `func`, `result` and `expected`
`os_type`	`Ubuntu`
`related_apps`	Applications involved
`source`	`Cognyzer`
`proxy`, `fixed_ip`	Both `false`; no task needs a proxy or a pinned address
`possibility_of_env_change`	`low` or `medium`
`model_pass_rate`	Empty; fill in from your own rollouts

Grading

Each evaluator runs check_include_exclude over the stdout of one or more commands in the VM. A run passes when every include marker is present and no exclude marker is. Markers are emitted by the checker scripts, so a failure tells you which specific requirement was missed rather than just that the task failed.

Several tasks go further than checking the output file:

audacity recovers a per segment frequency ladder from the exported audio to confirm the correct source window survived the trim, and scans shell history so an ffmpeg or sox shortcut cannot pass.
gimp and pinta correlate the produced image against a baseline regenerated at grading time, so an image that is merely the right size does not pass.
libreoffice-impress checks that no extra file whose name starts with the deck name was created anywhere on the machine.
libreoffice-writer recomputes the ODF section protection key from the requested password, which distinguishes the right password from any password.

Requirements

Each task names the VM image it expects in snapshot, and assumes that image already ships the application plus Python 3. Setup never installs packages and never reaches a package repository, so a run is reproducible as long as the image is.

Every file a task downloads at setup or grading time is served from assets/ in this repository. There are no third party download dependencies, so the tasks keep working regardless of what happens to any upstream cache.

No hints or solution notes are shipped with these tasks, and model_pass_rate is intentionally empty. Nothing here has been tuned against a particular model.

Downloads last month: 36

Total file size:

9.2 MB