Skip to main content

​
Reward Function Anatomy

This guide provides a detailed explanation of how reward functions are structured in the Reward Kit, focusing on the @reward_function decorator and the components that make up a complete reward function.

​
The @reward_function Decorator

The @reward_function decorator is the core mechanism that transforms a regular Python function into a reward function that can be used for evaluation and deployment. 
from reward_kit import reward_function, EvaluateResult, MetricResult

@reward_function
def my_reward_function(messages, original_messages=None, **kwargs):
    # Your evaluation logic here
    score = 0.75 # Example score
    reason = "Overall evaluation reason for my_reward_function"
    metrics_dict = {"example_metric": MetricResult(score=score, success=True, reason="Metric reason")}
    return EvaluateResult(score=score, reason=reason, metrics=metrics_dict)

​
What the Decorator Does

The @reward_function decorator performs several important functions:
  1. Input Validation: Ensures the function receives the expected parameters
  2. Output Standardization: Ensures the function returns a properly formatted EvaluateResult object
  3. Deployment Capability: Adds a .deploy() method to the function for easy deployment
  4. Backward Compatibility: Handles legacy return formats (tuples of score and metrics)

​
Under the Hood

Internally, the decorator wraps your function with logic that:
  1. Processes the input parameters
  2. Calls your function with the standardized inputs
  3. Handles any exceptions that occur during execution
  4. Formats the output as an EvaluateResult object
  5. Provides deployment capabilities through the .deploy() method

​
Function Parameters

A standard reward function has these parameters: 
from typing import List, Dict, Optional
from reward_kit import EvaluateResult

def reward_function(
    messages: List[Dict[str, str]],
    original_messages: Optional[List[Dict[str, str]]] = None,
    **kwargs
) -> EvaluateResult:
    # ...
    pass

​
Required Parameters

  • messages: A list of message dictionaries in the conversation, where each message has at least "role" and "content" keys. The last message is typically the one being evaluated.

​
Optional Parameters

  • original_messages: The conversation context, usually messages before the response being evaluated. If not provided, it defaults to messages[:-1].
  • **kwargs: Additional parameters that can be used to customize the evaluation.

​
Return Value

A reward function must return an EvaluateResult object: 
from reward_kit import EvaluateResult, MetricResult

# score, clarity_score, accuracy_score would be calculated by your logic
clarity_score = 0.8
accuracy_score = 0.7
final_score = 0.75

return EvaluateResult(
    score=final_score,  # Overall score between 0.0 and 1.0
    reason="Overall evaluation based on clarity and accuracy.",
    metrics={    # Component metrics
        "clarity": MetricResult(
            score=clarity_score,
            success=clarity_score >= 0.7,
            reason="The response clearly explains the concept"
        ),
        "accuracy": MetricResult(
            score=accuracy_score,
            success=accuracy_score >= 0.6,
            reason="Contains one minor factual error"
        )
    }
)

​
EvaluateResult Structure

  • score: The final aggregate score (typically between 0.0 and 1.0).
  • reason: An optional top-level explanation for the overall score.
  • metrics: A dictionary of component metrics (MetricResult objects), each with its own score, success flag, and explanation.
  • error: An optional string field to convey errors during evaluation.

​
Multi-Component Reward Functions

Complex reward functions often evaluate multiple aspects of a response: 
from reward_kit import reward_function, EvaluateResult, MetricResult
from typing import List, Dict, Optional

# Assume evaluate_clarity and evaluate_accuracy are defined elsewhere
def evaluate_clarity(response: str) -> float: return 0.8
def evaluate_accuracy(response: str) -> float: return 0.6


@reward_function
def comprehensive_evaluation(
    messages: List[Dict[str, str]],
    original_messages: Optional[List[Dict[str, str]]] = None,
    **kwargs
) -> EvaluateResult:
    response = messages[-1]["content"]
    metrics = {}

    # Evaluate clarity
    clarity_score = evaluate_clarity(response)
    metrics["clarity"] = MetricResult(
        score=clarity_score,
        success=clarity_score >= 0.7,
        reason=f"Clarity score: {clarity_score:.2f}"
    )

    # Evaluate accuracy
    accuracy_score = evaluate_accuracy(response)
    metrics["accuracy"] = MetricResult(
        score=accuracy_score,
        success=accuracy_score >= 0.6,
        reason=f"Accuracy score: {accuracy_score:.2f}"
    )

    # Combine scores (weighted average)
    final_score = clarity_score * 0.4 + accuracy_score * 0.6

    return EvaluateResult(score=final_score, reason="Comprehensive evaluation complete.", metrics=metrics)

​
Deployment Capabilities

The @reward_function decorator adds a .deploy() method to your function: 
# Assume my_reward_function is defined as above
# Deploy the function to Fireworks
evaluation_id = my_reward_function.deploy(
    name="my-evaluator",
    description="Evaluates responses based on custom criteria",
    force=True  # Overwrite if already exists
)

​
Deploy Method Parameters

  • name: ID for the deployed evaluator (required)
  • description: Human-readable description (optional)
  • force: Whether to overwrite an existing evaluator with the same name (optional)
  • providers: List of model providers to use for evaluation (optional)

​
Error Handling

Robust reward functions include proper error handling: 
from reward_kit import reward_function, EvaluateResult, MetricResult
from typing import List, Dict, Optional

@reward_function
def safe_evaluation(
    messages: List[Dict[str, str]],
    original_messages: Optional[List[Dict[str, str]]] = None,
    **kwargs
) -> EvaluateResult:
    try:
        # Ensure we have a valid response to evaluate
        if not messages or messages[-1].get("role") != "assistant":
            return EvaluateResult(
                score=0.0,
                reason="No assistant response found.",
                metrics={"error": MetricResult(
                    score=0.0,
                    success=False,
                    reason="No assistant response found"
                )}
            )

        # Your evaluation logic here
        # ...
        # For example:
        calculated_score = 0.0 # Placeholder for actual logic
        if calculated_score == 0 : raise ValueError("Simulated error")
        return EvaluateResult(score=1.0, reason="Successful evaluation", metrics={})


    except Exception as e:
        # Handle any unexpected errors
        return EvaluateResult(
            score=0.0,
            reason=f"Evaluation error: {str(e)}",
            metrics={"error": MetricResult(
                score=0.0,
                success=False,
                reason=f"Evaluation error: {str(e)}"
            )}
        )

​
Working with Metadata

You can pass additional configuration through the **kwargs parameter, often via a metadata dictionary. 
from reward_kit import reward_function, EvaluateResult, MetricResult
from typing import List, Dict, Optional, Any

# Assume base_score and metrics are calculated based on messages
def calculate_base_score_and_metrics(response_content: str, min_length: int) -> tuple[float, dict]:
    # Dummy implementation
    current_length = len(response_content)
    score = 1.0 if current_length >= min_length else 0.5
    return score, {"length_check": MetricResult(score=score, success=current_length >= min_length, reason=f"Length {current_length} vs min {min_length}")}

@reward_function
def configurable_evaluation(
    messages: List[Dict[str, str]], # Added type hints
    original_messages: Optional[List[Dict[str, str]]] = None, # Added type hints
    metadata: Optional[Dict[str, Any]] = None,
    **kwargs
) -> EvaluateResult:
    """Reward function that supports configuration via metadata."""
    metadata = metadata or {}
    response_content = messages[-1].get("content", "")

    # Get configurable thresholds from metadata
    min_length = metadata.get("min_length", 50)
    max_score_cap = metadata.get("max_score_cap", 1.0) # Renamed to avoid conflict with 'score'
    weight_factor = metadata.get("weight_factor", 1.0)

    # Use these parameters in your evaluation
    base_score, metrics = calculate_base_score_and_metrics(response_content, min_length)

    # Apply any metadata-based adjustments to the final score
    final_score = base_score * weight_factor
    final_score = min(final_score, max_score_cap) # Cap the score

    return EvaluateResult(score=final_score, reason="Configurable evaluation complete.", metrics=metrics)
When calling the function, you can pass this metadata: 
# Assume test_messages is defined
# result = configurable_evaluation(
#     messages=test_messages,
#     metadata={"min_length": 100, "weight_factor": 1.2}
# )

​
Next Steps

Now that you understand the structure of reward functions:
  1. Learn about the Core Data Types used in reward functions
  2. Explore Evaluation Workflows for testing and deployment
  3. See Code Examples for practical implementations
⌘I