This page is generated by joining eval games from a number of runs with 62241 extra games played between models of different runs and calculating ratings over all those games.

100 models were choosen from each recent run, 75 good models chosen evenly along the run and 25 of the strongest models.
Any games played between these models were kept and additional games have been played between models not in the same run, to provide a more accurate rating.

These ratings are less inflated than self-play elo because models face a wider more diverse set of opponents.