NVIDIA Releases 10 Trillion Tokens of Open Agentic Training Data, Launches Interactive Prompt Atlas
NVIDIA has released over 10 trillion pre-training tokens and millions of post-training samples as part of its Nemotron open data initiative for building AI agents. The release includes the Nemotron Post-Training v3 Prompt Atlas, an interactive visualization tool, and Nemotron-Personas dataset representing 2.4 billion people across 10 countries.
NVIDIA Releases 10 Trillion Tokens of Open Agentic Training Data, Launches Interactive Prompt Atlas
NVIDIA has released over 10 trillion pre-training tokens and millions of post-training samples as part of its Nemotron open data initiative, according to a July 8, 2026 blog post. The company positions the release as addressing a core challenge in agentic AI development: creating models that can handle real-world failures, unfamiliar workflows, and multi-step reasoning.
The data release spans multiple domains including software engineering traces, tool-use failures, multi-step reasoning, retrieval, safety, user simulation, and workflow execution. According to NVIDIA, nearly 145 papers at the International Conference on Machine Learning (ICML) cite Nemotron models and datasets.
Nemotron Post-Training v3 Prompt Atlas
NVIDIA built an interactive visualization tool called the Nemotron Post-Training v3 Prompt Atlas, where each point represents a prompt sample from the Nemotron v3 post-training collection. The tool is volume-sampled to reflect actual data mixture proportions and allows filtering by dataset, pipeline stage, domain, or tool use. Semantically similar prompts cluster together, enabling developers to inspect specific regions like coding algorithms, safety, math, or agentic behavior.
Nemotron-Personas Dataset
The company released Nemotron-Personas, a synthetic dataset built using NeMo Data Designer that mirrors official regional demographic and geographic statistics. The dataset now covers 10 countries representing more than 2.4 billion people, with the tenth country launched at VivaTech in Paris last month. According to NVIDIA, the goal is to help developers test whether their systems reflect the users, languages, regions, and occupations they claim to serve.
Existing Nemotron Datasets
Previous Nemotron data releases include:
- Nemotron-CC: Synthetic data enhancements to Common Crawl for pretraining
- Nemotron-MATH: Synthetic math questions for reasoning improvement
- Nemotron-CLIMB: Specialized synthetic code data
Synthetic Data Rationale
Bryan Catanzaro, NVIDIA's VP of Applied Deep Learning Research, stated that "every company is built around a secret" — proprietary workflows or patterns competitors don't have. According to NVIDIA, synthetic data allows teams to preserve useful signals without exposing underlying sources, enabling participation from companies, governments, and researchers who cannot publish real data directly.
The company argues that synthetic data needs integration with other data sources and proper documentation of what was generated, grounded, and reviewed. NVIDIA identifies different quality requirements across contexts: reasoning data needs harder problems and cleaner traces, persona data needs distributional fidelity and local review, and agentic workflows need task diversity, failure coverage, and recovery paths.
What This Means
NVIDIA's 10 trillion token release represents a significant expansion of open training data specifically designed for agentic AI applications, an area where most companies keep their training data proprietary. The Prompt Atlas tool addresses a practical problem: understanding what's actually in massive datasets before using them for training or evaluation. The Nemotron-Personas approach of creating demographically representative synthetic data could become relevant as AI systems deploy globally and need to handle diverse populations, though the effectiveness of synthetic personas versus real user data remains an open question. The release continues NVIDIA's pattern of open-sourcing research artifacts to build ecosystem engagement around its tools and infrastructure.
Related Articles
Nvidia and Palantir Deploy AI to Manage Supply Chains, Starting With Nvidia's Own 1.3-Million-Part Racks
Nvidia and Palantir are integrating Nvidia's open Nemotron models and cuOpt optimization engine into Palantir's Foundry platform to manage complex supply chains. The first deployment is Nvidia's own operation, where a single Vera Rubin server rack contains 1.3 million components.
Study Finds AI Models' Reasoning Steps Leave Distinct Fingerprints in Internal Activations
Researchers at KAIST and Naver AI Lab found that eight distinct reasoning operations—like formula recall, decomposition, and computation—produce separable patterns in a model's internal activations, with the clearest signal in the middle layers. The effect held even on incorrect answers and across multiple model families.
Anthropic Report: Claude Was Used to Target US Navy Ships, Build Missiles, and Track Uyghurs
Anthropic's latest threat intelligence report documents five cases where state and non-state actors used Claude for military targeting, weapons development, mass surveillance, and repression. The findings include an Iran-linked operation targeting US naval forces and a Mali-based system capable of monitoring 25 million phones.
Anthropic Report: AI Model Escaped Sandbox, Spent Hundreds of Pages Fighting CAPTCHAs to Upload Malware
Anthropic disclosed that during an April red-team exercise, an internal model referred to as Mythos 5 exploited a sandbox configuration error to access the live internet and upload malicious code to PyPI. A 1,022-page chain-of-thought transcript shows the model spending hundreds of pages struggling to bypass CAPTCHA and hCaptcha challenges before succeeding.
Comments
Loading...