NVIDIA GR00T: Breaking News

NVIDIA Shares Open Datasets on Hugging Face to Tackle AI Data Bottlenecks

Key Facts

• NVIDIA has released more than 2 petabytes of AI-ready training data across more than 180 datasets alongside 650+ open models.
• Datasets are published on Hugging Face with permissive licenses, training recipes, and evaluation frameworks available on GitHub.
• Key releases include the Physical AI Collection with 500K+ robotics trajectories and 1,700 hours of autonomous vehicle data spanning 25 countries.
• Nemotron Personas Collection provides millions of synthetic personas for sovereign AI, including 21M for India and 6M each for the US, Japan, and Brazil.
• La Proteina offers 455,000 synthetic atomistic protein structures for biology and drug discovery, developed with Oxford, Mila, and CIFAR researchers.

NVIDIA is addressing one of AI development's biggest challenges by openly sharing high-quality datasets on Hugging Face, the company announced in a collaborative blog post. The initiative aims to reduce the time and cost of building trustworthy AI systems and agents by making data, training recipes, and evaluation tools freely accessible to developers worldwide. By publishing permissively licensed datasets alongside its open models, NVIDIA seeks to accelerate progress in robotics, sovereign AI, biology, and other domains while promoting greater transparency in training data.

AI-Data Bottlenecks

High-quality dataset creation remains a major friction point in AI pipelines. Organizations frequently spend millions of dollars and many months — sometimes over a year — collecting, annotating, and validating data before training even begins. Domain expertise and robust evaluation frameworks continue to be scarce resources even after deployment.

NVIDIA is tackling this by publishing datasets on Hugging Face that developers can immediately use and build upon, paired with training recipes and evaluation frameworks on GitHub. To date, the company has shared more than 2 petabytes of AI-ready training data across more than 180 datasets and more than 650 open models.

Real-World Open Datasets

The releases span multiple domains critical to advancing AI. The Physical AI Collection targets robotics and autonomous systems with more than 500,000 robotics trajectories, 57 million grasps, and 15 terabytes of multimodal data. This data supported development of NVIDIA's GR00T reasoning vision-language-action model across various gripper types and sensor configurations. The collection has been downloaded more than 10 million times, with companies like Runway using it to develop its GWM-Robotics world model and Lightwheel leveraging it to refine robotics policies.

The dataset also includes one of the most geographically diverse autonomous vehicle collections available, featuring more than 1,700 hours of multi-sensor data from 7-camera configurations plus LiDAR and radar. The data covers 25 countries and over 2,500 cities, enabling perception benchmarking across varied driving environments.

Nemotron Personas and Biology Advances

The Nemotron Personas Collection consists of fully synthetic persona datasets grounded in real-world demographic distributions. These produce culturally authentic, diverse individuals across regions and languages at scale to support sovereign AI initiatives. Current population-scale datasets include 6 million personas for the United States, 6 million for Japan, 21 million for India, 6 million for Brazil (developed with WideLabs), and 888,000 for Singapore (developed with AI Singapore).

Real-world deployments demonstrate the value of these datasets. 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 enhance legal QA accuracy from 15.3% to 79.3% while reducing attack success rates from 7% to 0%. The personas also helped develop NVIDIA Nemotron-Nano-9B-v2-Japanese, which reached the top of the Nejumi leaderboard.

In biology, La Proteina provides a fully synthetic, atomistic protein dataset with 455,000 structures. It delivers a 73% structural diversity boost over prior baselines and offers design-ready molecular representations without privacy or licensing constraints. The dataset resulted from open collaboration with researchers from Oxford, Mila, and CIFAR.

SPEED-Bench and Broader Ecosystem

NVIDIA also released SPEED-Bench, a standardized benchmark for evaluating speculative decoding performance. It includes a Qualitative Split maximizing semantic diversity across 11 text categories and a Throughput Split organized by input sequence length.

The blog post emphasizes that open data access enables faster, more cost-effective model development while making evaluation and improvement easier across the ecosystem. This approach aligns with NVIDIA's broader strategy of releasing open models, tools, and techniques alongside data.

Impact

For developers and researchers, these releases lower barriers to building specialized AI systems. Teams can now access production-grade datasets instead of starting from scratch, potentially saving significant time and resources. The permissive licensing and accompanying tools enable immediate experimentation and customization for domain-specific applications.

The datasets particularly benefit sovereign AI efforts by providing culturally appropriate training data at scale. In robotics and physical AI, the multimodal collections accelerate development of capable agents that can reason, plan, and act safely in real-world environments. Biology researchers gain access to diverse molecular data without traditional constraints.

What's Next

NVIDIA indicates it is "just getting started" with open data releases. The company continues to build datasets across additional domains while expanding its portfolio of open models and tools. Future releases are expected to further address data bottlenecks as agentic systems become more prevalent and require increasingly sophisticated training data.

The collaborative approach with partners like Hugging Face demonstrates an industry shift toward shared data resources. As AI systems grow more autonomous, transparent and high-quality training data will play an even larger role in determining model behavior, reasoning capabilities, and safety.

Sources

• How NVIDIA Builds Open Data for AI
• NVIDIA Blog on Open Models, Data and Tools
• NVIDIA Launches Open Models and Data to Accelerate AI Innovation

All technical specifications, pricing, and benchmark data in this article are sourced directly from official announcements. Competitor comparisons use publicly available data at time of publication. We update our coverage as new information becomes available.