MediumNLPnlp-155

Contradiction Probability

Problem

Return the softmax probability of the contradiction class (index 1) over the NLI logits.

Use the max-subtraction trick for numerical stability. The logits must contain at least two entries.

Hint

Compute the full softmax first, then read the probability at index 1.

case 1

in: [[2,1,0]]

expected: 0.24472847105479764

case 2

in: [[0,0,0]]

expected: 0.3333333333333333

case 3

in: [[0,5,0]]

expected: 0.986703291042268

case 4

in: [[1,1]]

expected: 0.5

Starter codePython

import math
def contradiction_probability(logits):
    # Your code here
    pass

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