EvSys

HarborTask

One RL task - instruction + verifier spec.

Same shape whether the task came from a hand-written eval set, an HF dataset, or a generated rollout corpus. Runners materialize the prompt by feeding instruction to the policy, generate a rollout, and score it using verifier (whichever variant - see VerifierPayload).

Attributes

attributetask_idstr
attributeinstructionstr
attributeverifierVerifierPayload
attributemetadatadict
= field(default_factory=dict)

Functions

func__init__(self, task_id, instruction, verifier, metadata=dict()) -> None
paramself
paramtask_idstr
paraminstructionstr
paramverifierVerifierPayload
parammetadatadict
= dict()

Returns

None

On this page