NVIDIA Releases Massive Open Datasets for AI on Hugging Face
Key Facts
- NVIDIA has published more than 2 petabytes of permissively licensed AI-ready training data across over 180 datasets on Hugging Face.
- Releases include the Physical AI Collection with 500K+ robotics trajectories, 57M grasps, and 15TB of multimodal data used to develop the GR00T model.
- Nemotron Personas Collection offers synthetic persona datasets totaling nearly 40 million entries across the US, Japan, India, Brazil, and Singapore.
- La Proteina provides 455,000 synthetic atomistic protein structures with 73% greater structural diversity for biology and drug discovery.
- SPEED-Bench offers a standardized benchmark for evaluating speculative decoding performance with qualitative and throughput splits.
- All datasets are accompanied by training recipes and evaluation frameworks on GitHub.
Lead paragraph
NVIDIA is addressing one of AI development’s biggest bottlenecks by openly sharing more than 2 petabytes of high-quality training data on Hugging Face. In a new blog post published March 10, 2026, the company detailed its collaborative strategy to scale trustworthy AI systems through open datasets spanning robotics, autonomous vehicles, sovereign AI, biology, and evaluation benchmarks. Working with Hugging Face, NVIDIA aims to give developers immediate access to permissively licensed data, training recipes, and evaluation tools, reducing the time and cost of building specialized models and agentic systems.
The Data Bottleneck Challenge
AI progress depends as much on data quality as on model architecture or compute power. Organizations routinely spend millions of dollars and many months — sometimes over a year — collecting, annotating, and validating datasets before training even begins. Even after deployment, access to domain-specific evaluation frameworks and expert knowledge remains difficult.
NVIDIA’s strategy is to publish open datasets alongside its models and tools. According to the company’s announcement on the Hugging Face blog, this approach accelerates development while making evaluation and continuous improvement easier across the broader AI ecosystem. To date, NVIDIA says it has shared more than 2 petabytes of AI-ready training data through more than 180 datasets and over 650 open models.
Physical AI and Robotics Datasets
One of the most substantial releases is the Physical AI Collection. It includes more than 500,000 robotics trajectories, 57 million grasps, and 15 terabytes of multimodal data. These assets were used in developing NVIDIA’s GR00T reasoning vision-language-action model across multiple gripper types and sensor configurations.
The collection has seen strong adoption, with more than 10 million downloads. Companies are already building on the data: Runway used the open GR00T dataset to develop its recently released GWM-Robotics world model, while robotics simulation company Lightwheel is leveraging it to refine robotics policies.
The Physical AI Collection also features one of the most geographically diverse autonomous vehicle datasets available. It contains more than 1,700 hours of multi-sensor data captured with 7-camera configurations plus LiDAR and radar. The data spans 25 countries and over 2,500 cities, offering broad commercial usability that complements more limited academic datasets.
Nemotron Personas for Sovereign AI
The Nemotron Personas Collection provides fully synthetic persona datasets grounded in real-world demographic distributions. These generate culturally authentic, diverse individuals across regions and languages at population scale. Current datasets include:
- United States: 6 million personas
- Japan: 6 million personas
- India: 21 million personas
- Brazil: 6 million personas (developed with WideLabs)
- Singapore: 888,000 personas (developed with AI Singapore)
These synthetic personas are already delivering measurable results in production deployments. CrowdStrike used 2 million personas to improve natural language to CQL translation accuracy from 50.7% to 90.4%. In Japan, NTT Data and APTO leveraged the datasets to bootstrap domain-specific intelligence, raising legal QA accuracy from 15.3% to 79.3% while reducing attack success rates from 7% to 0%.
The personas also supported development of NVIDIA Nemotron-Nano-9B-v2-Japanese, which reached the top of the Nejumi leaderboard as a state-of-the-art sub-10B model.
Biology and Benchmark Contributions
In the biology domain, NVIDIA released La Proteina — a fully synthetic, atomistic protein dataset containing 455,000 structures. The dataset delivers a 73% boost in structural diversity over prior baselines and provides design-ready molecular representations without privacy or licensing constraints. The work resulted from open collaboration with researchers from Oxford, Mila, and CIFAR.
NVIDIA also introduced SPEED-Bench, a standardized benchmark for evaluating speculative decoding performance. It features two splits: a Qualitative Split that maximizes semantic diversity across 11 text categories, and a Throughput Split organized by input sequence length buckets ranging from 1K to 32K tokens.
Broader Ecosystem Strategy
NVIDIA’s open data initiative aligns with its wider push into open models and tools. The company has consistently released datasets alongside models, training techniques, and evaluation frameworks. This comprehensive approach supports the development of agentic AI systems that require high-quality, trustworthy training data to operate safely and effectively.
By hosting datasets on Hugging Face and providing training recipes on GitHub, NVIDIA lowers barriers for developers worldwide. The permissive licensing enables immediate use and modification, accelerating innovation across industries including robotics, healthcare, autonomous systems, and enterprise AI.
Impact on Developers and Industry
For developers, these releases provide production-ready data that would otherwise require enormous investment to create. Teams can now fine-tune or train specialized models using datasets already validated by NVIDIA’s research and engineering organizations.
The impact is particularly significant for sovereign AI initiatives. The Nemotron Personas datasets enable countries and organizations to develop culturally appropriate models with minimal reliance on proprietary or sensitive real-world data.
In robotics and physical AI, the open datasets are helping accelerate progress toward more capable and generalizable systems. The widespread adoption of the GR00T dataset demonstrates how open data can become foundational infrastructure for the entire ecosystem.
What's Next
NVIDIA indicated that the current releases represent only the beginning of its open data efforts. The company plans to continue expanding its portfolio of datasets across additional domains while improving the accompanying tools and evaluation frameworks.
As agentic and multimodal systems become more prevalent, high-quality open data will likely play an increasingly central role in determining which organizations can build competitive AI capabilities quickly and responsibly. NVIDIA’s strategy positions it as both a provider of cutting-edge hardware and a key contributor to the open data foundations that future AI systems will depend on.
The collaboration with Hugging Face makes these resources immediately accessible to millions of developers in the platform’s community, potentially amplifying their impact across research and commercial applications.