perplexity-from-loss — Microsoft case

Microsoft-inspired language-model evaluation service reports perplexity from the average token loss. You need to convert the supplied loss into perplexity accurately so model runs can be compared on the same scale.

Convert the mean negative log-likelihood per token (natural logarithm) into perplexity: $e^{\text{loss}}e^\{\backslash text\{loss\}\}$.

Implement solve(mean_nll).

Returns. Return the perplexity as a Python float.

Examples

Example 1

Input:

solve(2.0)

Output:

7.389056

Example 2

Input:

solve(0.0)

Output:

1.0