average reward · training rollouts

1 2 3 4 5 0 20 40 60 80 100 step
smoothed (ema 0.6)raw

average reward · held-out prompts

3.5 3.75 4 4.25 4.5 0 20 40 60 80 100 step
recorded value at each eval

eval reward is the mean over the held-out prompts at each checkpoint.