Building AI agents that can handle real-world chaos requires more than clever model weights. According to NVIDIA, the missing piece is high-quality open data, and the company is betting on synthetic data to provide it at scale.
More Than Model Weights
NVIDIA's Nemotron open data effort rests on a simple observation: an agent that cannot recover from a broken API call or an unfamiliar workflow is not really an agent. It is just an autocomplete tool with extra functions. Bridging that gap demands diverse data covering software engineering traces, tool-use failures, multi-step reasoning, retrieval, safety checks, user simulations, workflow execution, and eventually physical-world interactions.
At the recent International Conference on Machine Learning (ICML), nearly 145 papers cited Nemotron models and datasets. The company highlighted how synthetic data powers several key datasets:
- Nemotron-CC used synthetic data to improve the Common Crawl dataset for pretraining.
- Nemotron-MATH includes synthetic math questions to sharpen reasoning.
- Nemotron-CLIMB features specialized synthetic code examples.
NVIDIA releases open datasets partly to learn alongside the community and broaden those applications. But for agents, weights alone are insufficient. Reproducibility also depends on the datasets, curation choices, training recipes, and evaluation methods behind a model. Agent behavior must be inspectable. If a model calls tools, executes workflows, retrieves information, and acts across systems, developers need to understand the data that shaped those behaviors. Open data makes that inspection possible, and synthetic data is a key part of the solution.
Keep It Like a Secret
Bryan Catanzaro, NVIDIA's Vice President of Applied Deep Learning Research, noted that every company is built around a secret: a workflow, corpus, or customer pattern that competitors lack. Those secrets make AI useful, but companies cannot casually expose them. Synthetic data offers a way to preserve useful signals without revealing the underlying sources.
Catanzaro also emphasized the importance of a diverse and participatory AI ecosystem where many kinds of companies, researchers, governments, and communities can contribute. This is not just a philosophical stance; it is a practical data claim. If every model learns from the same narrow pool of data, models will begin to feel identical. The problem is that the most valuable data often sits inside organizations that cannot or will not publish it directly. Everyone benefits from a richer shared data layer, but no one wants to be the first to give away what makes them special. Synthetic data released openly is one way to change that dynamic.
Exploring Agent Data
As part of the Nemotron open data release, NVIDIA has made available over 10 trillion pretraining tokens and millions of post-training samples across many domains and data formats. To help developers explore what is actually inside Nemotron post-training data, the team built the Nemotron Post-Training v3 Prompt Atlas: an interactive visual map where each point represents a prompt sample drawn from the Nemotron v3 post-training collection. The samples are volume-sampled to reflect the honest proportions of the data mixture.
Users can apply color overlays and filters to reorganize the map by dataset, pipeline stage, domain, or tool use. Because semantically similar prompts cluster together, developers can zoom into regions covering coding algorithms, safety, math, or agentic behavior, inspect representative examples, and use that signal to curate data, build evaluations, or understand why a model behaves as it does.
Stay updated
Get the day's AI and automation news in your inbox. No spam, unsubscribe anytime.
Viva La Persona
Agents also need to understand the people they are built to support, and that is where data quality becomes local rather than universal. A toxicity classifier trained on English internet data can miss hostile messages in Korean or Japanese, where aggression is often encoded in politeness levels rather than obvious vocabulary. The same signal appears different in different contexts.
Nemotron-Personas attempts to address this by using locally grounded synthetic personas that capture the diversity and complexity of populations. Built with NeMo Data Designer, NVIDIA's compound-AI tooling for synthetic data generation, Nemotron-Personas mirrors official regional demographic and geographic statistics. The goal is not to recreate real people but to help developers test whether their systems reflect the users, languages, regions, and occupations they claim to serve. At VivaTech in Paris last month, NVIDIA launched the tenth country in the collection, which now represents more than 2.4 billion people.
When quality is local, only people who know that locality can build it: regional researchers, native speakers, subject-matter experts, and stakeholders who can inspect and correct alongside the developers. That is learning in public, not releasing data in isolation but building it collaboratively.
Ground Truths
Synthetic data must be integrated as part of a system of data sources. There are tradeoffs. It can reduce risk but does not remove the need for grounding, lineage, curation, evaluation, and human judgment.
NVIDIA suggests thinking in terms of synthetic thresholds: points where data can no longer be treated as purely real. That line is not always obvious. Real workflows, human feedback, model-generated traces, simulated users, and synthetic labels can all become intertwined. The answer is not to pretend synthetic data is fake or harmless. It is to document what was generated, what was grounded, what was reviewed, and what the data is meant to test. As more AI systems are trained on artificial information, the field needs better shared habits for inspecting, documenting, and debating these technologies in public.
Quality also means different things in different contexts. Reasoning data needs harder problems and cleaner traces. Persona data needs distributional fidelity and local review. Agentic workflows need task diversity, failure coverage, and recovery paths. The field is still more craft than formula.
That is why open methods matter. Synthetic data is not just about generating more examples. It is about asking better questions and making it possible for parties that otherwise could not sit at the same table: companies without giving away their secrets, governments without compromising privacy, and researchers without waiting for permission that may never come.
The scarce resource in AI is not tokens. It is trust between organizations. Synthetic data is one of the few tools available for building that trust.
NVIDIA hosted a livestream on Tuesday, July 7, 2026 titled "Why Open Data Matters" with a panel of experts. The Nemotron data collections are available on Hugging Face, along with the Prompt Atlas interactive map.