trinity.common.workflows.math_ruler_workflow module

Math workflow with RULER.

class trinity.common.workflows.math_ruler_workflow.MathRULERWorkflow(*, task: Task, model: ModelWrapper, auxiliary_models: List[ModelWrapper] | None = None)

基类：SimpleWorkflow

A workflow for math with RULER reward function.

Modified from MathWorkflow. Adapted from OpenPipe/ART

__init__(*, task: Task, model: ModelWrapper, auxiliary_models: List[ModelWrapper] | None = None)

reset(task: Task)

Note that in this workflow, MathRewardFn is only used for calculating the 'golden reward', whereasa the rewards used by RL training are calculated by RULER.

run() → List[Experience]

Modified from SimpleWorkflow.run

get_ruler_scores(responses: List[Experience], judger: Any) → Tuple[bool, List[float]]

Get RULER scores

class trinity.common.workflows.math_ruler_workflow.AsyncMathRULERWorkflow(*, task: Task, model: ModelWrapper, auxiliary_models: List[ModelWrapper] | None = None)

基类：MathRULERWorkflow

is_async: bool = True

async run_async() → List[Experience]

Modified from SimpleWorkflow.run