NVIDIA has released the Nemotron Post-Training v3 Prompt Atlas, an interactive visual map designed to help developers explore and understand the synthetic data used in its open models. The tool allows users to filter and zoom into clusters of prompt samples by dataset, pipeline stage, domain, or tool use, making agent behavior inspectable.
The atlas is built from Nemotron v3 post-training collection data, volume-sampled to reflect honest proportions of the data mixture. It aims to address the difficulty of building AI agents that can recover from broken API calls and unseen workflows by providing transparency into the training data.
NVIDIA also highlighted Nemotron-Personas, a dataset of locally grounded synthetic personas built using NeMo Data Designer to mirror regional demographics for testing agent inclusivity.