← vraj

models

my own scores for new models, three days after launch.

leaderboard

models, one row each
loading data/

harness bench

mean score per model and harness, the best cell in each row marked
loading data/

harness bench v2 is eight held-out tasks, one seed per task. the models land within 0.7 points of each other, which is inside one variant of noise: a smoke pass, not a ranking.

analyse