Surprisingly good low-slop creative writing benchmarks
#2
by GeoMaciolek - opened
The "Caliper Bench" leaderboard rank this model as among the lowest "slop" out there - really impressive, especially considering the size. I'm curious how that played out!
Thank you! I saw that, it was kind of unexpected. It might be because the dataset was based on my own personal writing, although it was expanded with AI. I did try to make sure the dataset was clean of overt sloppiness, like em dashes and annoying "not x, y." cliches. It was a pretty small dataset too, less than 500 rows, all single turn, but relatively high quality.