NVIDIA and Hugging Face have released an open dataset containing over 1.5 million instruction-trajectory pairs designed to improve AI agent reasoning and tool usage, according to a blog post from the companies. The release is part of NVIDIA's broader Nemotron open data initiative, which also includes more than 10 trillion pre-training tokens and synthetic datasets covering over 2.4 billion people across 10 countries.
Prompt Atlas for navigating the data
Alongside the dataset, NVIDIA introduced the Nemotron Post-Training v3 Prompt Atlas, an interactive visual map of prompt samples from the Nemotron v3 post-training mixture. The tool uses volume sampling to reflect domain proportions and lets developers filter prompts by pipeline stage, dataset, domain, or tool-use pattern, making it easier to inspect specific failure modes or multi-step reasoning traces.
Key figures from the release
- 1.5 million agent trajectories: instruction-trajectory pairs for training agents on tool execution, multi-step reasoning, and recovery paths.
- 10+ trillion pre-training tokens: including Nemotron-CC and Nemotron-CC-MATH, aimed at general and mathematical reasoning.
- 2.4 billion population coverage: the Nemotron-Personas dataset spans 10 countries and was built using NeMo Data Designer to match official regional demographic statistics.
- According to NVIDIA, Nemotron open models and datasets have already been cited in 145 ICML papers.
Personas and synthetic-data limitations
A key part of the release is Nemotron-Personas, intended to ground models in local demographics, occupations, and cultural context. A derivative dataset, Privasis, applies privacy-preserving synthetic records to legal, medical, and financial contexts in the United States.
NVIDIA researchers note that synthetic data cannot fully substitute for real-world grounding, lineage verification, and human review, and recommend that teams define explicit "synthetic thresholds" documenting which data was synthetically generated versus human-reviewed.
The combined release gives developers access to large-scale trajectory and demographic data for building and evaluating agentic models without needing proprietary data pools.