Task Definitions

Task definitions sample semantic parameters once, realize an initial layout, and render a concrete Task_Spec for one episode. Parameters and geometry use independent random streams. Save the resulting tasks to repeat an evaluation with the same instructions, layouts, and scoring questions.

TaskParameters

TaskParameters()

Typed semantic values; subclasses enforce their task's domain constraints.

Methods

record() -> dict[str, Task_ParameterValue]

P

P = TypeVar('P', bound=TaskParameters)

TaskDefinition

class TaskDefinition(ABC, Generic[P])

Authoring API; only the concrete result of sample goes to the station.

Attributes

name: str

Canonical task ID, matching the documentation and sampled Task_Spec.name.

layout: LayoutSpec

max_rollout_time: float

Methods

sample_parameters(random: Random) -> P

Draw a valid assignment, including distinctness and order constraints.

task(parameters: P, max_rollout_time: float) -> Task_Spec

Render setup, policy instruction, and scoring from the same assignment.

sample(seed: int, *, max_rollout_time: float | None = None, parameters: P | None = None, max_attempts: int = 10000) -> Task_Spec

Realize an episode. An explicit parameter assignment overrides its random draw.

The definition supplies the time limit unless explicitly overridden. Geometry retries never change the semantic assignment.

sample_set(seed: int, *, episodes: int, max_rollout_time: float | None = None) -> tuple[Task_Spec, ...]

Generate a reproducible episode set to reuse across policies.