reward-kit

0.4.1 verified Mon Apr 27 auth: no python

A Python library for defining, testing, and using reward functions, particularly for RLHF and AI alignment. Version 0.4.1, pre-1.0, active development with weekly releases.

pip install reward-kit

Common errors

error ModuleNotFoundError: No module named 'reward_kit' ↓

cause Library not installed or installed in wrong environment.

fix

Run 'pip install reward-kit' in the correct Python environment.

error AttributeError: module 'reward_kit' has no attribute 'evaluate_reward' ↓

cause Import path changed in version 0.4.0.

fix

Use 'from reward_kit import evaluate_reward' instead of 'from reward_kit.utils import evaluate_reward'.

error ValueError: Unknown model: 'morality' ↓

cause Model name 'morality' is deprecated in 0.4.0+ and removed in some versions.

fix

Use 'morality_v2' or import Model enum: reward_kit.Model.MORALITY_V2.

Warnings

breaking In version 0.4.0, the evaluate_reward function was moved from reward_kit.utils to top-level reward_kit. Old imports will break. ↓

fix Use 'from reward_kit import evaluate_reward' instead of 'from reward_kit.utils import evaluate_reward'

breaking The RewardFunction class constructor 'model' parameter changed from string to enum in 0.4.1. Strings still work but are deprecated. ↓

fix Pass a Model enum value, e.g., reward_kit.Model.MORALITY, or keep string but expect future removal.

deprecated The 'morality' model name is deprecated in favor of 'morality_v2' as of 0.4.0. The v1 model will be removed in 0.5. ↓

fix Use model='morality_v2' or reward_kit.Model.MORALITY_V2.

gotcha By default, reward functions require an internet connection to call remote APIs. Offline mode is not yet supported. ↓

fix Ensure you have internet access or expect ConnectionError.

Imports

RewardFunction

wrong

from reward_kit.core import RewardFunction

correct

from reward_kit import RewardFunction

In version 0.3+ RewardFunction moved to top-level

evaluate_reward

wrong

from reward_kit.utils import evaluate_reward

correct

from reward_kit import evaluate_reward

API changed in 0.4.0

Quickstart

Initialize a reward function, load a dataset, and evaluate a sample.

from reward_kit import RewardFunction
from reward_kit.data import MoralityQA
import os

# Use your own API key or prompt key
api_key = os.environ.get('REWARD_KIT_API_KEY', 'your-key-here')

reward_fn = RewardFunction(model='morality', api_key=api_key)
dataset = MoralityQA()
score = reward_fn.evaluate(dataset[0])
print(score)