Bits per Character from Perplexity

Problem

Convert perplexity measured per token into bits per character: log2(perplexity) * num_tokens / num_chars. This normalizes across tokenizers so models with different vocabularies are comparable. Return 0.0 when perplexity <= 0, num_tokens < 0, or num_chars <= 0.

Test cases

case 1

in: [16,8,16]

expected: 2

case 2

in: [2,4,8]

expected: 0.5

case 3

in: [0,5,5]

expected: 0

case 4

in: [1024,10,10]

expected: 10

Starter codePython

import math
def bits_per_char(perplexity, num_tokens, num_chars):
    # Your code here
    pass

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