Reward Model Ensemble Variance

Problem

Compute the population variance of reward predictions across an ensemble: the mean of (p - mean)^2 over the predictions. An empty ensemble returns 0.0.

Hint

High ensemble variance marks preference pairs the reward model is unsure about.

Test cases

case 1

in: [[1,2,3]]

expected: 0.6666666666666666

case 2

in: [[5,5,5]]

expected: 0

case 3

in: [[]]

expected: 0

case 4

in: [[-1,1]]

expected: 1

Starter codePython

def reward_model_ensemble_variance(predictions):
    # Your code here
    pass

4 test cases · run your code against them in the editor