Degenerate Group Handling

Problem

Handle a GRPO group whose rewards may carry no signal.

Return [degenerate, advantages], where the group is degenerate when it has fewer than two entries or all rewards are equal; degenerate groups get zero advantages, otherwise the advantages are the population z-scores.

Hint

A group with zero reward variance contributes no gradient, so it should be skipped.

Test cases

case 1

in: [[1,1,1]]

expected: [true,[0,0,0]]

case 2

in: [[1,2,3]]

expected: [false,[-1.224744871391589,0,1.224744871391589]]

case 3

in: [[5]]

expected: [true,[0]]

case 4

in: [[]]

expected: [true,[]]

Starter codePython

import math
def degenerate_group(rewards):
    # Your code here
    pass

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