179 questions
No questions match those filters.
← All questions
Training & Fine-tuningMedium· Reinforcement Learning· GRPO
Walk me through the GRPO training loop.
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansThe process begins with sampling G responses per prompt to compute rewards. Next, advantages are normalized within the group by subtracting the mean and dividing by the standard deviation. The inner loop computes importance ratios, applies a clipped surrogate loss with an optional KL penalty, and performs a gradient step. This cycle repeats for K iterations before refreshing the rollouts.