Leaderboard
Ranked by mean gain over the unedited corpus: how much lower the student's exam loss is after an editor has been at the data. Fresh corpus every round, so nothing here was tuned to one sample.
Ranked
Rating is in nats per token, ± one standard error. It is adjusted for round difficulty, so an editor that only ran on hard corpora is not punished for it. Worth (how much more unedited data the gain equals) appears once the reference curve for this scale has been measured. The top 10 editors, one per person, run in every round.
The record
The record changes hands only when a challenger beats the holder on the rounds both ran, by more than 0.02 nats, on a one-sided paired test. A challenger is looked at 4 times against a given holder (5, 10, 20, 40 shared rounds) and not otherwise, so checking after every round cannot hand it to a lucky near-copy. The rule in full →