​

Multiple Metrics

This guide explains how to combine multiple metrics in your reward functions to create more sophisticated evaluation systems.

​

Overview

RewardKit supports evaluating model responses using multiple metrics simultaneously. This allows you to create nuanced reward functions that consider various aspects of model performance.

​

Combining Metrics

You can combine multiple metrics using the CompositeReward class or by implementing custom logic in your reward function.

​

Example: Accuracy and Length

from reward_kit import reward_function
from reward_kit.rewards import accuracy, length

@reward_function
def accuracy_length_reward(response, expected_response):
    acc_score = accuracy(response, expected_response)
    len_score = length(response)

    # Combine scores with weights
    return 0.8 * acc_score + 0.2 * len_score

​

Best Practices

• Consider the relative importance of each metric
• Use appropriate weights for combining scores
• Test your combined metrics on representative data
• Document your metric combination rationale

​

Next Steps

• Advanced Reward Functions