Ask the Experts: Inside Nemotron Post-Training | Nemotron Labs

In this Ask the Experts session, NVIDIA AI researchers explain how they built the final Nemotron model checkpoints through post-training to enhance model intelligence and enable agentic capabilities. They’ll walk through the tools they used, how they structured the data pipeline, and the open resources available for the community to build on.

Learn how NeMo Data Designer was used to create task-specific synthetic datasets, NeMo Gym to build and manage training environments, and NeMo RL for reinforcement learning training. Learn about the post-training datasets published on HuggingFace and the recipes and weights on the Nemotron GitHub that are available under open licenses.

Different stages of post-training pipeline that goes into building SOTA models

How NVIDIA NeMo libraries was used to implement and run post-training workflows

Which datasets, recipes, and model weights are openly available, and where to find them