Group Size For Learning Signal

Problem

Find the smallest GRPO group size whose probability of containing both a success and a failure reaches threshold.

For group size g that probability is 1 - p ** g - (1 - p) ** g. Return -1 if no size up to max_group qualifies or if p is not strictly between 0 and 1.

Hint

Groups where every sample succeeds or every sample fails produce zero advantages.

Test cases

case 1

in: [0.5,0.8,4]

expected: 4

case 2

in: [0.5,0.7,4]

expected: 3

case 3

in: [0.1,0.5,4]

expected: -1

case 4

in: [0,0.5,8]

expected: -1

case 5

in: [1,0.5,8]

expected: -1

Starter codePython

def group_size_for_signal(p, threshold, max_group):
    # Your code here
    pass

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