math

BasePointWiseReward

Bases: BaseReward

Point-wise reward module for individual response evaluation.

Evaluates each response independently without considering relative ranking.

Source code in rm_gallery/core/reward/base.py

class BasePointWiseReward(BaseReward):
    """
    Point-wise reward module for individual response evaluation.

    Evaluates each response independently without considering relative ranking.
    """

    @abstractmethod
    def _evaluate(
        self, sample: DataSample, **kwargs
    ) -> RewardResult[RewardDimensionWithScore]:
        """
        Processes a single response to generate reward metrics.

        Parameters:
            sample (DataSample): Single-response data sample
            **kwargs: Evaluation parameters

        Returns:
            RewardResult[RewardDimensionWithScore]: Response-specific reward metrics
        """
        ...

    def _parallel(
        self,
        func: Callable,
        sample: DataSample,
        thread_pool: ThreadPoolExecutor | None = None,
        **kwargs,
    ) -> DataSample:
        """
        Processes responses in a data sample using parallel or sequential execution.

        This method applies the provided function to each response in the sample,
        either in parallel using a thread pool or sequentially. Results are merged
        back into the corresponding response objects.

        Parameters:
            func (Callable): Function to apply to each response. Should accept a
                DataSample and return an object with 'details' and 'extra_data' attributes.
            sample (DataSample): Input sample containing multiple responses to process
            thread_pool (ThreadPoolExecutor | None): Optional thread pool for parallel execution
            **kwargs: Additional arguments passed to func

        Returns:
            DataSample: Modified copy of input sample with reward metrics updated in each response

        The method creates a deep copy of the input sample to avoid modifying original data.
        When using a thread pool, it submits tasks for each response and waits for completion
        before merging results. Response objects are updated with both reward details and
        additional metadata from processing results.
        """
        sample = sample.model_copy(deep=True)
        futures = []
        for i, output in enumerate(sample.output):
            # Create sub-sample for individual response processing
            subsample = DataSample(
                unique_id=sample.unique_id, input=sample.input, output=[output]
            )

            if thread_pool:
                futures.append(
                    (
                        i,
                        thread_pool.submit(
                            func, sample=subsample, thread_pool=thread_pool, **kwargs
                        ),
                    )
                )
            else:
                result = func(
                    sample=subsample,
                    thread_pool=thread_pool,
                    **kwargs,
                )
                output.answer.reward.details += result.details
                output.answer.additional_kwargs[self.name] = result.extra_data

        # Process parallel execution results
        if thread_pool:
            wait([future[-1] for future in futures], return_when=ALL_COMPLETED)
            # Merge results back into sample outputs
            for i, future in futures:
                result = future.result()
                output = sample.output[i]
                output.answer.reward.details += result.details
                output.answer.additional_kwargs[self.name] = result.extra_data

        for output in sample.output:
            if len(output.answer.reward.details) > 0:
                output.answer.reward.score = sum(
                    r.score for r in output.answer.reward.details
                ) / len(output.answer.reward.details)

        return sample

DataSample

Bases: BaseModel

Complete data sample structure for reward modeling training and evaluation.

Represents a single interaction with input context, multiple possible outputs, and associated metadata for comprehensive reward model training.

Attributes:

Name	Type	Description
unique_id	str	Unique identifier for tracking and deduplication
input	List[ChatMessage]	Conversation context as list of chat messages
output	List[DataOutput]	List of possible responses with evaluations
task_category	Optional[str]	Optional categorization for task-specific analysis
source	Optional[str]	Origin dataset or system that generated this sample
created_at	datetime	Timestamp for temporal tracking
metadata	Optional[Dict]	Additional context and debugging information

Source code in rm_gallery/core/data/schema.py

class DataSample(BaseModel):
    """
    Complete data sample structure for reward modeling training and evaluation.

    Represents a single interaction with input context, multiple possible outputs,
    and associated metadata for comprehensive reward model training.

    Attributes:
        unique_id: Unique identifier for tracking and deduplication
        input: Conversation context as list of chat messages
        output: List of possible responses with evaluations
        task_category: Optional categorization for task-specific analysis
        source: Origin dataset or system that generated this sample
        created_at: Timestamp for temporal tracking
        metadata: Additional context and debugging information
    """

    unique_id: str = Field(..., description="Unique identifier for the data")
    input: List[ChatMessage] = Field(default_factory=list, description="input")
    output: List[DataOutput] = Field(default_factory=list, description="output")
    task_category: Optional[str] = Field(default=None, description="task category")
    source: Optional[str] = Field(default=None, description="source")
    created_at: datetime = Field(default_factory=datetime.now, description="createdAt")
    metadata: Optional[Dict] = Field(default=None, description="metadata")

    def update(self, sample: "DataSample") -> "DataSample":
        """
        Merge another sample's data into this sample for combining evaluations.

        Updates additional_kwargs and reward details from the source sample
        while preserving the original structure.

        Args:
            sample: Source sample to merge data from

        Returns:
            Self with updated data for method chaining
        """
        self.input[-1].additional_kwargs.update(sample.input[-1].additional_kwargs)
        for i, output in enumerate(self.output):
            output.answer.additional_kwargs.update(
                sample.output[i].answer.additional_kwargs
            )
            output.answer.reward.details.extend(sample.output[i].answer.reward.details)

            if output.steps:
                for j, step in output.steps:
                    step.additional_kwargs.update(
                        sample.output[i].steps[j].additional_kwargs
                    )
                    step.reward.details.extend(sample.output[i].steps[j].reward.details)
        return self

    class Config:
        arbitrary_types_allowed = True
        json_encoders = {datetime: lambda v: v.isoformat()}

update(sample)

Merge another sample's data into this sample for combining evaluations.

Updates additional_kwargs and reward details from the source sample while preserving the original structure.

Parameters:

Name	Type	Description	Default
sample	DataSample	Source sample to merge data from	required

Returns:

Type	Description
DataSample	Self with updated data for method chaining

Source code in rm_gallery/core/data/schema.py

def update(self, sample: "DataSample") -> "DataSample":
    """
    Merge another sample's data into this sample for combining evaluations.

    Updates additional_kwargs and reward details from the source sample
    while preserving the original structure.

    Args:
        sample: Source sample to merge data from

    Returns:
        Self with updated data for method chaining
    """
    self.input[-1].additional_kwargs.update(sample.input[-1].additional_kwargs)
    for i, output in enumerate(self.output):
        output.answer.additional_kwargs.update(
            sample.output[i].answer.additional_kwargs
        )
        output.answer.reward.details.extend(sample.output[i].answer.reward.details)

        if output.steps:
            for j, step in output.steps:
                step.additional_kwargs.update(
                    sample.output[i].steps[j].additional_kwargs
                )
                step.reward.details.extend(sample.output[i].steps[j].reward.details)
    return self

MathVerifyReward

Bases: BasePointWiseReward

Verifies mathematical expressions using the math_verify library, supporting both LaTeX and plain expressions

Source code in rm_gallery/gallery/rm/math/math.py

@RewardRegistry.register("math_verify_reward")
class MathVerifyReward(BasePointWiseReward):
    """
    Verifies mathematical expressions using the math_verify library, supporting both LaTeX and plain expressions
    """

    name: str = Field(default="math_verify", description="Math verification reward")
    timeout_score: float = Field(default=0.0, description="Score to assign on timeout")

    def _evaluate(
        self, sample: DataSample, **kwargs
    ) -> RewardResult[RewardDimensionWithScore]:
        """
        Verify mathematical expressions

        Args:
            sample: Data sample containing mathematical content

        Returns:
            RewardResult: Reward result containing verification score
        """
        generated = sample.output[0].answer.content.strip()
        reference = sample.output[0].answer.label.get("reference", "").strip()

        score = 0.0
        reason = "Verification failed or timed out"

        try:
            # Parse the reference (gold) answer
            # Use both LatexExtractionConfig and ExprExtractionConfig for maximum flexibility
            gold_parsed = parse(
                reference,
                extraction_config=[LatexExtractionConfig(), ExprExtractionConfig()],
            )

            # Parse the generated answer
            pred_parsed = parse(
                generated,
                extraction_config=[LatexExtractionConfig(), ExprExtractionConfig()],
            )

            # If both parsing succeeded and we have results
            if gold_parsed and pred_parsed:
                # Use the first parsed result from each
                gold_expr = gold_parsed[0]
                pred_expr = pred_parsed[0]

                # Verify if they match
                if verify(gold_expr, pred_expr):
                    score = 1.0
                    reason = f"({gold_parsed}, {pred_parsed})"
                else:
                    score = 0.0
                    reason = f"({gold_parsed}, {pred_parsed})"
            else:
                score = 0.0
                reason = f"Parsing failed - gold: {gold_parsed}, pred: {pred_parsed}"

        except Exception as e:
            score = self.timeout_score
            reason = f"Exception occurred: {str(e)}"

        return RewardResult(
            name=self.name,
            details=[
                RewardDimensionWithScore(
                    name=self.name,
                    score=score,
                    reason=str(reason),
                )
            ],
            extra_data={
                "generated": generated,
                "reference": reference,
                "score": score,
            },
        )

RewardDimensionWithScore

Bases: RewardDimension

Pointwise/Stepwise reward dimension with a numerical score.

Attributes:

Name	Type	Description
score	float	Numerical value representing the reward magnitude

Source code in rm_gallery/core/reward/schema.py

class RewardDimensionWithScore(RewardDimension):
    """
    Pointwise/Stepwise reward dimension with a numerical score.

    Attributes:
        score (float): Numerical value representing the reward magnitude
    """

    score: float = Field(default=..., description="score")

RewardRegistry

A registry management system for reward modules that maps module names to their corresponding implementation classes.

This class provides a centralized repository for registering and retrieving reward modules by string identifiers. Modules can be registered using decorators and later accessed by their string identifiers.

Attributes:

Name	Type	Description
_registry	Dict[str, Type[BaseReward]]	Internal dictionary storing the mapping between reward module names and their classes.

Source code in rm_gallery/core/reward/registry.py

class RewardRegistry:
    """A registry management system for reward modules that maps module names to their corresponding implementation classes.

    This class provides a centralized repository for registering and retrieving reward modules by string identifiers.
    Modules can be registered using decorators and later accessed by their string identifiers.

    Attributes:
        _registry: Internal dictionary storing the mapping between reward module names and their classes.
    """

    # Dictionary mapping reward module names to their corresponding classes
    _registry: Dict[str, Type[BaseReward]] = {}

    @classmethod
    def register(cls, name: str):
        """Create a decorator to register a reward module class with a specified identifier.

        The decorator pattern allows classes to be registered while maintaining their original identity.

        Args:
            name: Unique string identifier for the reward module
            module: The BaseReward subclass to be registered

        Returns:
            A decorator function that registers the module when applied to a class
        """

        def _register(module: Type[BaseReward]):
            """Internal registration function that stores the module in the registry.

            Args:
                module: The BaseReward subclass to be registered

            Returns:
                The original module class (unchanged)
            """
            cls._registry[name] = module
            return module

        return _register

    @classmethod
    def get(cls, name: str) -> Type[BaseReward] | None:
        """Retrieve a registered reward module class by its identifier.

        Provides safe access to registered modules without raising errors for missing entries.

        Args:
            name: String identifier of the reward module to retrieve

        Returns:
            The corresponding BaseReward subclass if found, None otherwise
        """
        assert name in cls._registry, f"Reward module '{name}' not found"
        return cls._registry.get(name, None)

    @classmethod
    def list(cls) -> str:
        """
        Returns:
            A list of all registered reward modules
        """
        info = []
        for name, module in cls._registry.items():
            info.append(
                pd.Series(
                    {
                        "Name": name,
                        "Class": module.__name__,
                        "Scenario": module.__doc__.strip(),
                    }
                )
            )

        info_df = pd.concat(info, axis=1).T
        # info_str = info_df.to_markdown(index=False)
        info_str = tabulate(
            info_df,
            headers="keys",
            tablefmt="grid",
            maxcolwidths=[50] * (len(info_df.columns) + 1),
            # showindex=False,
        )
        # info_str = tabulate(info_df, headers='keys', tablefmt='github')
        return info_str

get(name) classmethod

Retrieve a registered reward module class by its identifier.

Provides safe access to registered modules without raising errors for missing entries.

Parameters:

Name	Type	Description	Default
name	str	String identifier of the reward module to retrieve	required

Returns:

Type	Description
Type[BaseReward] | None	The corresponding BaseReward subclass if found, None otherwise

Source code in rm_gallery/core/reward/registry.py

@classmethod
def get(cls, name: str) -> Type[BaseReward] | None:
    """Retrieve a registered reward module class by its identifier.

    Provides safe access to registered modules without raising errors for missing entries.

    Args:
        name: String identifier of the reward module to retrieve

    Returns:
        The corresponding BaseReward subclass if found, None otherwise
    """
    assert name in cls._registry, f"Reward module '{name}' not found"
    return cls._registry.get(name, None)

list() classmethod

Returns:

Type	Description
str	A list of all registered reward modules

Source code in rm_gallery/core/reward/registry.py

@classmethod
def list(cls) -> str:
    """
    Returns:
        A list of all registered reward modules
    """
    info = []
    for name, module in cls._registry.items():
        info.append(
            pd.Series(
                {
                    "Name": name,
                    "Class": module.__name__,
                    "Scenario": module.__doc__.strip(),
                }
            )
        )

    info_df = pd.concat(info, axis=1).T
    # info_str = info_df.to_markdown(index=False)
    info_str = tabulate(
        info_df,
        headers="keys",
        tablefmt="grid",
        maxcolwidths=[50] * (len(info_df.columns) + 1),
        # showindex=False,
    )
    # info_str = tabulate(info_df, headers='keys', tablefmt='github')
    return info_str

register(name) classmethod

Create a decorator to register a reward module class with a specified identifier.

The decorator pattern allows classes to be registered while maintaining their original identity.

Parameters:

Name	Type	Description	Default
name	str	Unique string identifier for the reward module	required
module		The BaseReward subclass to be registered	required

Returns:

Type	Description
	A decorator function that registers the module when applied to a class

Source code in rm_gallery/core/reward/registry.py

@classmethod
def register(cls, name: str):
    """Create a decorator to register a reward module class with a specified identifier.

    The decorator pattern allows classes to be registered while maintaining their original identity.

    Args:
        name: Unique string identifier for the reward module
        module: The BaseReward subclass to be registered

    Returns:
        A decorator function that registers the module when applied to a class
    """

    def _register(module: Type[BaseReward]):
        """Internal registration function that stores the module in the registry.

        Args:
            module: The BaseReward subclass to be registered

        Returns:
            The original module class (unchanged)
        """
        cls._registry[name] = module
        return module

    return _register

RewardResult

Bases: BaseModel, Generic[T]

Container for reward calculation results with generic type support.

Attributes:

Name	Type	Description
name	str	Identifier of the reward module that generated this result
details	List[T]	Collection of detailed reward information items
extra_data	dict	Additional metadata or context information

Source code in rm_gallery/core/reward/schema.py

class RewardResult(BaseModel, Generic[T]):
    """
    Container for reward calculation results with generic type support.

    Attributes:
        name (str): Identifier of the reward module that generated this result
        details (List[T]): Collection of detailed reward information items
        extra_data (dict): Additional metadata or context information
    """

    name: str = Field(default=..., description="reward module name")
    details: List[T] = Field(default_factory=list, description="reward details")
    extra_data: dict = Field(default_factory=dict, description="extra data")