The reward before the filter read higher
Mean reward per step over every answer the engine finished, for 12 steps with seed 1. Without the filter: 0.657 on average. With the filter, counted before it dropped any group: 0.686.
0.5
0.6
0.7
0.8
0
3
6
9
11
step
mean reward of all finished answers
no filter
DAPO filter, before the filter
Mean over 12 steps: 0.657 without the filter, 0.686 with it. One seed.