Types

A credentialed account and a signed agreement are needed, for example a PhysioNet DUA.

Anyone can fetch the data; no account or agreement is needed.

Access is granted case by case, for example an AWS-IAM grant or a direct request.

Optional confidence for this particular application.

Optional human-readable description of the annotation.

Name identifying the annotation.

Optional metadata that is part of the reusable content.

The annotation's public name.

Optional metadata for this particular application.

Optional per-instance provenance (the rater, method, or model behind this value). Unlike description, which the descriptor holds once per key, source can differ between annotations that share a key.

Where on the Record timeline this annotation applies. Its owner determines its scope. None means it is record-scoped context with no place in time.

Optional physical unit of value. Give a unit string (for example "years") or a :class:pint.Unit. Construction validates the unit against the shared registry. An unrecognized string raises ValueError. Construction stores a pint.Unit as its canonical name.

A native scalar or a text handle. Call read() for string and string-list values.

None makes it a pure marker, which needs a span to mark something.

Which of the three annotation shapes this key uses.

Optional human-readable description of the annotation.

Name of the annotation this descriptor projects.

Optional physical unit of the value.

Manifest value-type tag (bool, int, float, str, list, or map).

Annotations describing this task itself.

Tasks from which this task was derived.

The task's id, unique among the tasks of its dataset.

:meth:~timenet.dataset.TimeFDataset.add_task sets it. The reader sets it when it reads the task back.

Raises:

Existing annotation occurrences supplied to the model as context.

Optional override. None infers the input kinds when resolved or written.

Records supplied to the model, in semantic order.

Optional JSON-compatible task metadata.

Call read() for the prompt; None identifies an unprompted task.

Call read() for the explanation; None means no rationale is present.

The validated override, or kinds inferred from the current inputs.

Inference reads Signal specs within the scope, the prompt, and text input annotations. It does not load values or inspect targets. NO_INPUT means there are no input Records. The result is not cached: connectors can assign inputs after construction.

Optional region of the inputs to which the task applies.

The partition the task belongs to: train, validation, or test. None when the dataset does not partition its tasks.

Existing annotation occurrences that form the expected output.

Ordered native values and stored objects that form the expected output.

Whether the value is held in memory, so read() needs no storage access.

Annotations describing this task itself.

Tasks from which this task was derived.

The task's id, unique among the tasks of its dataset.

:meth:~timenet.dataset.TimeFDataset.add_task sets it. The reader sets it when it reads the task back.

Raises:

Existing annotation occurrences supplied to the model as context.

Optional override. None infers the input kinds when resolved or written.

Records supplied to the model, in semantic order.

The allowed labels and whether a target can contain more than one label.

Optional JSON-compatible task metadata.

Call read() for the prompt; None identifies an unprompted task.

Call read() for the explanation; None means no rationale is present.

The validated override, or kinds inferred from the current inputs.

Inference reads Signal specs within the scope, the prompt, and text input annotations. It does not load values or inspect targets. NO_INPUT means there are no input Records. The result is not cached: connectors can assign inputs after construction.

Optional region of the inputs to which the task applies.

The partition the task belongs to: train, validation, or test. None when the dataset does not partition its tasks.

Existing annotation occurrences that form the expected output.

Ordered native values and stored objects that form the expected output.

How a user obtains the data (see :class:Access). OPEN needs nothing.

An import into a child dataset does not change it. The level that applies to a release with parents is :func:~timenet.types.access.effective_access.

Where to obtain access (the DUA or credentialing page). Required when access is not OPEN.

How to cite the dataset, when the source asks for attribution.

HuggingFace-style org/name pair, case-sensitive, exactly one slash.

Semantic version of the compiled TimeNet release.

Free-text description of the dataset.

Kinds of data the dataset contains.

Legal license of the source data, as an SPDX-style identifier.

Where to read the full license text. Required and nonblank when license is :attr:License.OTHER.

Human-readable display name.

Exact parent releases indexed by connector-local aliases.

Upstream release, tag, or commit, separate from the compiled dataset version.

Link to the dataset's origin, if any.

Free-form tags for search and grouping.

Version of the card's own field schema.

Type-level descriptors for the dataset's annotation keys.

Built-in Task subclasses the dataset declares, resolved from the registry.

Descriptors for the dataset's measurement modalities.

What people are doing: labeled human activities and daily behavior, usually from wearables.

Declare motion as well when the signals record the movement itself.

Recorded sound: speech, cough, and environmental audio.

The heart: ECG, heart rate, and heart sounds.

Economic activity: macroeconomic indicators, prices, sales, and demand.

Energy systems: electricity load, generation, power grids, and transformers.

The natural world: weather, climate, air quality, and hydrology.

Financial markets: asset prices, returns, trading volumes, and exchange rates.

No particular field: synthetic or demo data, or generic series without a stated origin.

A corpus that mixes such data with field-specific components declares those domains too.

Human physiology and clinical care: biosignals, vital signs, clinical records, glucose.

The umbrella for cardiology, respiratory, and sleep. Declare it alongside them.

Machines and industrial processes: engines, turbines, bearings, and production equipment.

Covers condition monitoring, degradation, fault diagnosis, and remaining useful life, including simulated run-to-failure data.

Movement of a body or an object: accelerometer, gyroscope, magnetometer, and motion capture.

Software and IT infrastructure: service metrics and cloud monitoring.

Breathing and the lungs: respiration, airflow, oxygen saturation, and cough.

Sleep studies: polysomnography and sleep staging.

Movement of people and goods through transport networks: road traffic, congestion, and ridership.

Telemetry from a vehicle's own machinery, such as engine health, belongs to industrial.

Annotations describing this task itself.

Tasks from which this task was derived.

The task's id, unique among the tasks of its dataset.

:meth:~timenet.dataset.TimeFDataset.add_task sets it. The reader sets it when it reads the task back.

Raises:

Existing annotation occurrences supplied to the model as context.

Optional override. None infers the input kinds when resolved or written.

Records supplied to the model, in semantic order.

Optional JSON-compatible task metadata.

Call read() for the prompt; None identifies an unprompted task.

Call read() for the explanation; None means no rationale is present.

The validated override, or kinds inferred from the current inputs.

Inference reads Signal specs within the scope, the prompt, and text input annotations. It does not load values or inspect targets. NO_INPUT means there are no input Records. The result is not cached: connectors can assign inputs after construction.

Optional region of the inputs to which the task applies.

The partition the task belongs to: train, validation, or test. None when the dataset does not partition its tasks.

Existing annotation occurrences that form the expected output.

Ordered native values and stored objects that form the expected output.

What the label means, such as The subject is awake..

The text that a target holds for this label.

The allowed labels in a fixed order.

If False, a target holds exactly one label from the set.

If True, a target holds all labels from the set that apply. An empty target means that no label applies.

The name of the set within a dataset, such as sleep_stage.

The name of each label, in label order.

A function that returns the text.

Always False.

The number of characters or list items.

If the caller omits length, the first access reads the text to calculate it.

A function that returns the text.

Always False.

The number of characters or list items.

If the caller omits length, the first access reads the text to calculate it.

SPDX's value for a source that states no license; its terms are unknown, not unrestricted.

Escape hatch for a license outside this list; requires a license_url on the card.

Annotations describing this task itself.

Tasks from which this task was derived.

The task's id, unique among the tasks of its dataset.

:meth:~timenet.dataset.TimeFDataset.add_task sets it. The reader sets it when it reads the task back.

Raises:

Existing annotation occurrences supplied to the model as context.

Optional override. None infers the input kinds when resolved or written.

Records supplied to the model, in semantic order.

Optional JSON-compatible task metadata.

Call read() for the prompt; None identifies an unprompted task.

Call read() for the explanation; None means no rationale is present.

The validated override, or kinds inferred from the current inputs.

Inference reads Signal specs within the scope, the prompt, and text input annotations. It does not load values or inspect targets. NO_INPUT means there are no input Records. The result is not cached: connectors can assign inputs after construction.

Optional region of the inputs to which the task applies.

The partition the task belongs to: train, validation, or test. None when the dataset does not partition its tasks.

Existing annotation occurrences that form the expected output.

Optional name of the predicted quantity.

Ordered native values and stored objects that form the expected output.

Optional physical unit shared by numerical target items.

The exclusive upper bound of the span, in its own frame.

The region is half-open, so a point ends one unit past its position and an interval ends at its stored end. Each concrete leaf supplies this.

Whether this span marks a single position rather than a bounded interval.

The range's own exclusive end, stop.

The frame its bounds read in, as stored on disk.

Whether this span marks a single position rather than a bounded interval.

How many steps the interval covers, the horizon h step-based forecasting libraries speak in.

The position, or the start of the interval, as a step ordinal from the series' first stored step.

Last step ordinal, exclusive.

The single series whose steps the span counts.

One step past the ordinal, so the half-open range covering it is [start, +1).

The frame its bounds read in, as stored on disk.

The position, or the start of the interval, as a step ordinal from the series' first stored step.

The single series whose steps the span counts.

The exclusive upper bound of the span, in its own frame.

The region is half-open, so a point ends one unit past its position and an interval ends at its stored end. Each concrete leaf supplies this.

The frame its bounds read in, as stored on disk.

Whether this span marks a single position rather than a bounded interval.

The position, or the start of the interval, as a step ordinal from the series' first stored step.

The single series whose steps the span counts.

The occurrences attached to this object.

Annotations describing this task itself.