Length Penalty Reward

Problem

Apply a length penalty to rewards: r_i - alpha * length_i for each pair of reward and length.

The signature is length_penalty_reward(rewards, lengths, alpha).

Hint

Penalizing length discourages padding and rambling generations.

case 1

in: [[1,2],[10,20],0.01]

expected: [0.9,1.8]

case 2

in: [[0],[5],0.1]

expected: [-0.5]

case 3

in: [[2,3],[0,4],0.5]

expected: [2,1]

case 4

in: [[1],[3],0]

expected: [1]

Starter codePython

def length_penalty_reward(rewards, lengths, alpha):
    # Your code here
    pass

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