Why open data is needed for agent AI and why it’s the way to scale synthetic data.
Image: Nemotron Post Training v3 Prompt Atlas
More than the weight of the model
Building AI agents is difficult because the real world does not behave like a benchmark.
An agent that can’t recover from a broken API call or an unfamiliar workflow isn’t really an agent at all. It is an autocompleter with tools. The transition from one to the other is a matter of data. That is, software engineering traces, tool usage failures, multi-step inference, acquisition, safety, user simulation, workflow execution, and ultimately physical world interaction. That’s where NVIDIA Nemotron’s open data products exist.
NVIDIA recently highlighted how open models are driving AI research and showing up at the popular International Conference on Machine Learning (ICML), where nearly 145 papers cited Nemotron models and datasets. Synthetic data plays an important role in that entire ecosystem.
Nemotron-CC used synthesis to enhance the popular Common Crawl dataset for pre-training. Nemotron-CC-MATH utilizes synthetic mathematics questions to improve reasoning. Nemotron Pretraining is an extensive collection spanning general, coded, mathematical, and synthetic data across trillions of tokens.
One of the reasons NVIDIA releases open datasets is to collaborate with the community and learn to extend these various applications.
Open weights are important. But for agents, weight is only part of the story. Reproducibility also depends on the dataset, curation choices, training recipes, and evaluation methods behind the model.
Agent behavior must be inspectable. As models call tools, execute workflows, retrieve information, and operate throughout the system, developers need to understand the data that shaped those behaviors. Open data makes it possible to inspect and explain agent behavior. Synthetic data is a key piece of the puzzle that makes that possible.
please keep it a secret
Brian Catanzaro, vice president of applied deep learning research at NVIDIA, recently said, “Every company is built around secrets.” This is a workflow, corpus, and customer patterns that our competitors don’t have. These secrets make AI useful, but companies should not reveal them casually. Synthetic data provides teams with a way to preserve useful signals without exposing the underlying source.
Brian also talks about fostering a diverse and participatory AI ecosystem where different types of companies, researchers, governments, and communities can contribute. It’s not just a value statement. This is a data request.
If all your models learn from the same narrow pool of data, you shouldn’t be surprised if your models start to feel the same. The challenge is that the most useful data is often located within an organization that cannot or will not be exposed directly. Everyone benefits from a richer shared data layer. No one wants to be the first to give away what makes them special.
Openly published synthetic data is one way to change that calculation.
Exploring agent data
As part of Nemotron Open Data, we have released over 10 trillion pre-training tokens and millions of post-training samples across many domains and data shapes. This is a lot to understand, and the raw dataset table isn’t very helpful.
We built the Nemotron Post-Training v3 Prompt Atlas to make it easy to explore what your Nemotron Post-Training data actually contains. This is an interactive visual map where each point is a prompt sample extracted from the Nemotron v3 post-training collection and volumetrically sampled to reflect an honest proportion of the data mix.
Color overlays and filters allow you to reorganize your map by dataset, pipeline stage, domain, or tool usage. Because semantically similar prompts are clustered together, you can zoom in on areas like coding algorithms, safety, mathematics, and agent behavior to examine representative examples and use that signal to curate data, build evaluations, and understand why your model behaves the way it does.
viva la persona
Agents also need to understand the people they are supporting, and this is where “data quality” becomes local rather than universal. A malicious classifier trained on English internet data may miss hostile messages in Korean or Japanese. In Korean and Japanese, aggression is often encoded at the level of politeness rather than overt vocabulary. Same signal, different context. The team is already grounding agents in this way.
Nemotron-Persona is one attempt to address this. Locally based synthetic personas capture the diversity and complexity of populations. Built using NeMo Data Designer, NVIDIA’s cutting-edge composite AI tool for synthetic data generation, the Nemotron-persona reflects official regional demographic and geographic statistics. The goal is not to recreate real people. In a way, this helps developers test whether their systems reflect the users, languages, regions, and professions they claim to serve. Last month at VivaTech in Paris, we announced the 10th country in our collection. Currently, its population exceeds 2.4 billion.
If quality is local, it can only be built by those who know its locality: local researchers, native speakers, subject matter experts, and stakeholders who can inspect and fix it together. It’s about learning in public and building data collaboratively rather than releasing it in isolation.
earthly truth
Synthetic data must be integrated as part of the data source system. There are trade-offs. Risk can be reduced, but it does not eliminate the need for evidence, pedigree, curation, evaluation, and human judgment.
One helpful way to think about this is through the use of a “synthesis threshold,” the point at which data can no longer be treated as purely real. The boundaries are not always clear. Actual workflows, human feedback, model-generated traces, simulated users, and synthetic labels can all intertwine. The answer is to not pretend that synthetic data is fake or harmless. Documenting what was generated, what was based on it, what was reviewed, and what the data is intended to test. As more AI systems are trained on artificial information, we will need better sharing habits for inspecting artificial information, documenting it, and discussing these technologies in public.
Quality has different meanings depending on the situation. Data inference requires harder problems and cleaner traces. Persona data requires distributional fidelity and local reviews. Agent workflows require task diversity, failure handling, and recovery paths. This field is still more of a craft than a formality.
That’s why an open approach is so important. Synthetic data does more than just generate more examples. It’s about asking better questions and allowing parties who might not otherwise be able to sit at the same table: companies without giving away secrets, governments without violating privacy, and researchers without waiting for permissions they may never get.
AI’s scarce resources are not tokens. It’s about trust between organizations. Synthetic data is one of the few tools to build it.
On Tuesday, July 7, 2026, we hosted a livestream on “Why Open Data Matters” with an amazing panel. It’s worth checking out in conjunction with Hugging Face’s Nemotron data collection.
Stay up to date on NVIDIA Nemotron by subscribing to NVIDIA News and following NVIDIA AI on the Nemotron channel on LinkedIn, X, YouTube, and Discord.
Access the open Nemotron model at Hugging Face and our collection of NIM microservices and developer samples at build.nvidia.com.
