Report
Astra reportedly used the same seed in every reinforcement-learning rollout
A user says they caught the issue at step seven, when rollouts were almost identical except for hardware nondeterminism.
TLDR
The user says Astra chose the same seed for every reinforcement-learning rollout for reproducibility. They say they spotted it at step seven, when the rollouts were almost identical apart from hardware nondeterminism.
Combined views
23.1K
5 Sources, first seen ago
