#pairwise-loss

A reward model is trained with the pairwise loss $-\log \sigma\big(r(x, y_w) - r(x, y_l)\big)$, where $y_w$ is the response the labeler preferred and $y_l$ the other one. On one comparison it currently outputs $r(x, y_w) = 0.5$ and $r(x, y_l) = 1.7$. What is the loss on this pair (natural log)?