trinity.common.rewards.rlcr_reward module

trinity.common.rewards.rlcr_reward module#

A safe port of the five-component RLCR reward for math tasks.

class trinity.common.rewards.rlcr_reward.RLCRRewardFn(weights: ~collections.abc.Mapping[str, ~typing.Any] | None = None, answer_parser: ~collections.abc.Callable[[str], ~typing.Any] = <function simple_answer_parser>, verifier: ~collections.abc.Callable[[~typing.Any, ~typing.Any], bool] = <function verify_with_timeout>)[source]#

Bases: RewardFn

Return already weighted RLCR format, accuracy, and calibration components.

__init__(weights: ~collections.abc.Mapping[str, ~typing.Any] | None = None, answer_parser: ~collections.abc.Callable[[str], ~typing.Any] = <function simple_answer_parser>, verifier: ~collections.abc.Callable[[~typing.Any, ~typing.Any], bool] = <function verify_with_timeout>) → None[source]#