Isn't the range of the JS divergence [0,1]? Why are there values greater than 1 in Figure 2?

voidism / DoLa

Official implementation for the paper "DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models"

https://arxiv.org/abs/2309.03883

419 stars 50 forks source link

Isn't the range of the JS divergence [0,1]? Why are there values greater than 1 in Figure 2? #10

Closed Jeryi-Sun closed 8 months ago

Jeryi-Sun commented 8 months ago

voidism commented 8 months ago

Hi,

The numbers in Figure 2 are actually scaled by $10^5$ so they are all between [0, 1]. We forgot to include this detail in the first version of our arxiv paper. But we have included this in the openreview version (See the caption of Figure 2): https://openreview.net/pdf?id=Th6NyL07na