#dpo

5-stage loop design and safety mechanisms for self-hosted SLMs to autonomously learn and improve from production traces based on Karpathy's autoresearch concept

Production configuration for running NeMo-RL (GRPO) and TRL (DPO) training jobs with labeled preference datasets on Karpenter Spot node pools and Volcano Gang Scheduling.

EKS-based 5-stage pipeline that automatically promotes Langfuse traces to training data and connects GRPO/DPO preference tuning with Canary deployment.