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/folder_based_builder/folder_based_builder.py", line 251, in _split_generators
raise ValueError(
"`file_name`, `*_file_name`, `file_names` or `*_file_names` must be present as dictionary key in metadata files"
)
ValueError: `file_name`, `*_file_name`, `file_names` or `*_file_names` must be present as dictionary key in metadata files
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.
ICLWorld
Code and execution instructions
ICLWorld evaluates whether agents can learn task-specific rules from video demonstrations or matched textual context and apply them in interactive Web, OS and Robot environments.
The dataset contains 160 tasks: 75 Web, 35 OS and 50 Robot. Each task has three rule variants, A/B/C. All domains share one metadata.jsonl, containing 480 task–variant records.
Documentation
| Document | Purpose |
|---|---|
README.md |
Dataset structure, field definitions, difficulty criteria and R/I/E definitions |
| TASK_REFERENCE_EN.md | Per-task scenarios, A/B/C rules, condition availability and video links for all 160 tasks |
Execution plans and launch instructions belong to the codebase. Experimental results are documented separately from the dataset.
Dataset structure
metadata.jsonl
asset/
cua/
iclworld_web_01/...
iclworld_os_01/...
rua/
iclworld_robot_01/...
environment/...
README.md
TASK_REFERENCE_EN.md
asset/cua/ contains Web/OS videos and required state data. asset/rua/ contains Robot videos and required physical environment resources, including geometry, textures and scene assets. The dataset has 330 Web/OS videos and 401 Robot videos: 150 real/sim demonstration pairs plus 101 condition-specific video inputs. All Robot videos have zero audio tracks. Real/sim identifies the demonstration modality; physical Robot execution uses simulation.
Task identifiers use iclworld_<domain>_<number>. Video names use the same identifier and a rule suffix, for example iclworld_web_01_a.mp4 and iclworld_robot_01_a_real.mp4. Condition-specific Robot videos append _r, _i or _e; an additional numerical suffix distinguishes multiple inputs when necessary. All resource paths are relative to the dataset root.
Difficulty
Difficulty describes the execution requirements of a task. A/B/C share the task's difficulty; the rule letters do not denote difficulty levels. Difficulty and the R/I/E condition are separate annotations.
| Domain | Easy | Medium | Hard |
|---|---|---|---|
| Web | Apply a learned rule to multiple objects or local decisions within one application. | Complete a dependent business workflow within a primary application, using the learned rule as a local module and combining relevant information into a usable deliverable. | Complete an end-to-end workflow across applications or business modules, transferring actual content and delivering it to the correct destination. |
| OS | Repeatedly apply one learned method to objects in a persistent desktop context, such as renaming, filing or formatting. | Complete a business goal with dependent steps in a primary desktop application; the learned method is one local module. | Complete a workflow across desktop applications, transferring real files, configurations or results while satisfying the final business goal. |
| Robot | Perform basic simple manipulation. | Perform multiple operations or coordinated dual-arm manipulation. | Perform fine manipulation, such as screw handling or card turning, or a long-horizon workflow that uses the learned rule as one module. |
| Domain | Easy | Medium | Hard | Total |
|---|---|---|---|---|
| Web | 25 | 25 | 25 | 75 |
| OS | 11 | 12 | 12 | 35 |
| Robot | 7 | 19 | 24 | 50 |
| Total | 43 | 56 | 61 | 160 |
Generalization
| Label | Meaning | Condition |
|---|---|---|
| R | Reproduction | Execute the task in the demonstrated original setting, using the demonstrated rule and original input/state. |
| I | Independent and identically distributed (IID) | For a fixed task and rule, demonstration and evaluation instances are independently sampled from the same task distribution. |
| E | Environment generalization | Execute under a changed environment while preserving the learned rule's semantics and task objective. |
I denotes the IID condition. A different instance can be sampled under I, but novelty of an input or object is not its definition. I does not designate a deliberate shift in the input or object distribution. The three labels apply consistently across Web, OS and Robot.
A task's generalization entries identify its available data conditions and condition-specific public prompts. Experiment subsets, random seeds and run schedules are specified in the codebase.
| Domain | R | I | E |
|---|---|---|---|
| Web | 25 | 75 | 75 |
| OS | 11 | 35 | 35 |
| Robot | 45 | 42 | 50 |
| Total task families | 81 | 152 | 160 |
Counts are distinct task families per supported condition; A/B/C variants and input modalities are not counted separately. For Web and OS, R replays the first example of the corresponding variant’s demonstration. I and E share the same instance, goal and evaluator when paired: E changes the Web workspace or OS desktop interface. For Robot, E pairs with I where I is supported and otherwise with R.
Metadata schema
Each line of metadata.jsonl is one JSON object for one task and one A/B/C variant. Every domain uses the same fields. Annotation text may be Chinese or English; the task reference provides English scenarios and rule descriptions.
| Field | Definition |
|---|---|
task_name |
Numbered task identifier, such as iclworld_web_01 |
task_description |
Scenario and task objective |
domain |
web, os or robot |
difficulty |
easy, medium or hard, as defined in Difficulty |
generalization |
Available R/I/E entries, as defined in Generalization |
variant |
A, B or C rule version |
videos |
Video records with names, relative paths and modality |
resource_dependencies |
Relative paths to required non-video data files or asset directories |
text_context |
Textual specification of the demonstrated rule |
prompt |
Public task objective without the hidden rule answer |
qa |
Rule-understanding question, options and evaluator answer |
Generalization entries
| Field | Definition |
|---|---|
condition |
R, I or E |
prompt |
Public task goal for this condition |
resource_dependencies |
Relative paths to the data resources required by this condition |
Video entries
| Field | Definition |
|---|---|
name |
Video filename |
path |
Video path relative to the dataset root |
type |
demonstration for Web/OS; real or sim for Robot |
generalization |
Optional R/I/E label for a condition-specific video input |
QA entries
| Field | Definition |
|---|---|
question |
Rule-understanding question |
options |
Mapping from option labels to option text |
correct_answer |
Correct option label, available only to the evaluator |
Data and execution boundaries
The dataset supplies task annotations, contexts, public prompts, QA and required data assets. The codebase supplies environment implementations, experiment subsets, seeds, model configuration, budgets, launchers, graders and Harbor workspace generation. Model logs and result tables are separate artifacts.
Text contexts are provided only under a declared text-input condition. Video-only and no-context agents must not read them. QA agents receive the question and options, never correct_answer. Initial states, resource bindings and grading information remain evaluator-side. The complete task reference contains rules and is also evaluator-side documentation.
- Downloads last month
- 120