Accounting for sampling parameters in prod RL
Score centering · 16 weight updates per redeployment · batch size 64 · three random seeds each
No top-p or top-k
Top-p 0.95 · top-k 20
0.2
0.3
0.4
0.5
0.6
0.7
0
48
96
144
192
Optimizer steps
Reward