Custom reward functions for multi-turn reinforcement learning with Amazon Nova Forge

Amazon Web Services · 2 months ago · 1

Amazon Nova Forge added multi-turn reinforcement fine-tuning support where you define a reward function that scores full episode rollouts inside your own environment via BYOO, or use a serverless option for multi-turn RL. The worked example trains Amazon Nova Lite 2.0 on 500 unique programming tasks using GRPO with LoRA on SageMaker HyperPod. Training outcomes can change based on reward design because GRPO learns only from within-group reward variation, so incorrect or collapsed reward components can prevent useful learning even when training curves look normal.