Command Palette
Search for a command to run...
Dataset Summary | NVIDIA Open Sources Nemotron Datasets: Over 10TB of Tokens + 40M Training Samples, Covering Mathematical Reasoning, Code Generation, and Multilingual dialogue.
Training data is becoming a key variable in the competition for large models. When the number of parameters is no longer the only barrier, the quality, structure, and task suitability of the data begin to determine the model's true performance in dimensions such as inference, code, and interaction.
NVIDIA's Nemotron dataset series is a data resource system built in response to this trend. It contains more than 10TB of tokens and 40M post-training samples, covering the entire training lifecycle from the base model to the agent workflow.Unlike traditional pre-training data that focuses solely on scale, the Nemotron series emphasizes targeted coverage of different capability dimensions, with each dataset corresponding to a specific aspect of model capability building.
This article compiles 15 Nemotron datasets, covering several core capabilities in current large-scale model training:General text pre-training, supervised fine-tuning (SFT), code generation, mathematical reasoning, and multilingual Persona dialogue data.It provides systematic data support for the model from basic capability building to task capability optimization.
The design philosophy behind these datasets reveals a clear shift: large-scale model training is moving from "training stronger models with more data" to "training more suitable models with more accurate data." The approach taken by the Nemotron series is a microcosm of this trend. This means that the ability to design training data is becoming a key variable determining the upper limit of a model's capabilities.
The following datasets have all been included in HyperAI (hyper.ai), and we hope to provide a systematic data resource reference for researchers and developers of large models.
More high-quality datasets:
Dataset Recommendation
1. Nemotron-CC-v2 pre-trained dataset
* Use online:
Nemotron-CC-v2 is a successor to Nemotron-CC released by NVIDIA in 2025. The related paper is titled "NVIDIA Nemotron Nano 2: An Accurate and Efficient Hybrid Mamba-Transformer Reasoning Model".
This dataset builds on the existing English web corpus by adding eight Common Crawl snapshots from 2024–2025, performing global deduplication and English filtering. It also uses Qwen3-30B-A3B to synthesize and restate web content, supplemented with Diverse Question Answering (Diverse QA), and further translated into 15 languages to enhance multilingual logical reasoning and general knowledge pre-training. Its significance lies in advancing the effective approach of "high-quality English webpages → synthesized Diverse QA" to a new level, combining updated web crawling and multilingual expansion into a systematic approach. Through rigorous deduplication, filtering, and reproducible distribution, it facilitates direct integration into various pre-training pipelines.
2. Nemotron-Pretraining-SFT Supervised Fine-Tuning Dataset
* Use online:
Nemotron-Pretraining-SFT-v1 is a synthetic generative dataset released by NVIDIA in 2025. The related paper is "NVIDIA Nemotron Nano 2: An Accurate and Efficient Hybrid Mamba-Transformer Reasoning Model", which aims to enhance the model's capabilities in tasks such as instruction following, reasoning, code and general question answering.
This dataset is designed for STEM, academic, logical reasoning, and multilingual scenarios. It is expanded from high-quality mathematical and scientific materials and combines graduate-level academic texts with finely tuned SFT data to construct complex multiple-choice and analytical questions (including complete solutions/approaches), covering various tasks such as mathematics, coding, general knowledge, and logical reasoning. In the official statistics of the Nemotron pre-training data, SFT-related categories (such as Math SFT, Code SFT, and General SFT) constitute a significant proportion, making it easy for users to filter the required subsets based on metadata for reproducible experiments.
3. Nemotron-Pretraining-Code dataset
* Use online:
Nemotron-Pretraining-Code-v1 is a set of carefully selected large-scale code datasets built on GitHub and released by NVIDIA in 2025. The related paper is titled "NVIDIA Nemotron Nano 2: An Accurate and Efficient Hybrid Mamba-Transformer Reasoning Model".
This dataset, filtered through multi-stage deduplication, license enforcement, and heuristic quality checks, contains LLM-generated code question-and-answer pairs for 11 programming languages. In addition to 175.1 B of high-quality synthesized code tokens, the data also includes metadata (approximately 747.4 B tokens) to facilitate user reproduction.
4. Nemotron-Pretraining-Code Pre-training Dataset
* Use online:
Nemotron-Pretraining-Code-v3 is one of the code pre-training datasets created by NVIDIA for the Nemotron 3 series of large language models. It aims to enhance the code understanding, generation, completion, and reasoning capabilities of LLM. This dataset belongs to the code corpus part of the NVIDIA Nemotron Pretraining Data series. The related paper is Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning.
This dataset contains a total of 146.3 million files. Data collection ended on September 30, 2025. This dataset is an incremental version, only including the source code files added to this series compared to previous versions. It needs to be used in conjunction with v1 and v2 to form a complete code pre-training corpus.
5. Nemotron-CC-Math Mathematical Pre-training Dataset
* Use online:
Nemotron-CC-Math is a high-quality, large-scale pre-trained dataset with a focus on mathematics, released by NVIDIA and Boston University in 2025. The related paper is titled "Nemotron-CC-Math: A 133 Billion-Token-Scale High Quality Math Pretraining Dataset". It aims to preserve and display high-value mathematical and code content, thereby driving the next wave of intelligent, globally capable language models.
This dataset contains 133 billion tokens and was built from Common Crawl using an extraction and normalization pipeline based on NVIDIA Lynx and a lightweight LLM. While preserving the equation and code format structure, it unifies the mathematical content into an editable LaTeX format, achieving, for the first time on a web scale, coverage of multiple (including long-tail) mathematical formats; its advantages have been validated in several benchmarks.
6. Nemotron-Personas-Korea (Korean synthetic human dataset)
* Use online:
Nemotron-Personas-Korea is a dataset of synthetic Korean characters released by NVIDIA in 2026, designed to comprehensively reflect the diversity and characteristics of the Korean population. This dataset is the first large-scale dataset of synthetic Korean characters, primarily used to expand the diversity of Sovereign AI synthetic data, mitigate data and model bias, and improve the diversity of model responses.
This dataset contains 1 million records, each with 7 virtual characters, totaling approximately 7 million character descriptions and about 1.7 billion tokens, of which 1 billion tokens are related to the characters. The data covers 17 metropolitan autonomous regions (시도) and 252 cities/districts (시군区) in South Korea, containing 209,167 unique names. This dataset was generated based on official information provided by the Korean Statistical Information Service (KOSIS), the Supreme Court of Korea, the National Health Insurance Corporation, the Korea Rural Economic Research Institute, and NAVER Cloud. All characters are completely virtual and do not correspond to any real personal identity information; data on minors under the age of 19 has also been excluded.
7. Nemotron Personas France (French synthetic person dataset)
* Use online:
Nemotron Personas France is a French synthetic character dataset released in 2026 by NVIDIA in collaboration with Pleias. It contains synthetic character data generated based on real-world French demographics, geography, and personality traits. The aim is to provide diverse synthetic character data to support model development by reflecting the geographical and demographic distribution of France.
This dataset contains 6,000,000 instances of French people, distributed across 1,000,000 records. Each record provides 22 fields (such as name, gender, age, marital status, occupation, etc.) and includes various person types (scholars, sports enthusiasts, art enthusiasts, food enthusiasts, and travel enthusiasts, etc.).
8.Nemotron-Personas-Brazil Brazilian Synthetic Character Dataset
* Use online:
Nemotron-Personas-Brazil is a synthetic character dataset for Brazil released in 2026 by NVIDIA in collaboration with WideLabs. It aims to showcase the diversity and richness of Brazil's population to more comprehensively reflect the multidimensional potential population distribution, including regional diversity (such as the North, Northeast, and Midwest), ethnic background, education level, and occupational distribution.
This dataset contains 1,000,000 records, each containing 6 composite characters. Each record includes 6 character fields and 14 context fields, which are statistically constructed based on Brazil's official population structure and labor market distribution. The data covers the geographical and demographic distribution of all 26 states and the Federal District of Brazil.
9.Nemotron Personas USA (USA) Personality Dataset
* Use online:
Nemotron-Personas-USA is a large-scale synthetic user profile dataset released by NVIDIA in 2025. It aims to support the training and evaluation of large language models (LLMs) and intelligent agent systems in tasks such as dialogue generation, persona simulation, user modeling, and diverse behavior analysis. The dataset contains approximately 1 million virtual person records, totaling 6 million persona fields and 16 contextual information fields. It covers all 50 states of the United States, as well as Puerto Rico and the Virgin Islands, encompassing 29,000 zip codes (ZCTAs) and 15,200 cities/regions, providing a relatively complete reflection of the geographical and social distribution of the US population.
The dataset contains approximately 970,000 unique names and covers more than 560 occupational categories. The occupational distribution references real-world occupational statistics, ensuring good social representativeness. Each data point consists of multidimensional fields, including structured demographic information such as age, gender, education level, income, occupation, and location, as well as natural language persona descriptions such as interests, values, lifestyle, and personal goals, forming a composite persona representation that combines structured information with unstructured text.
10. Nemotron-Personas-Japan: A dataset of Japanese synthetic human figures.
* Use online:
Nemotron-Personas-Japan is a dataset of synthetic human figures released by NVIDIA in 2025. It aims to showcase the diversity and richness of Japan's population and is primarily used to support the development of sovereign AI systems, training of large language models, and to reduce bias in synthetic data.
This dataset contains 1 million records, each with 6 virtual characters, totaling approximately 6 million character descriptions and about 1.4 billion tokens, of which 850 million tokens are related to the characters. The data covers all 47 prefectures of Japan, more than 1,500 occupational categories, and more than 950,000 unique names. The dataset was generated based on official Japanese demographic data, geographical distribution, and personality trait distribution. All characters are completely virtual and do not correspond to any real personal identity information. Data on minors under the age of 18 has also been excluded.
11. Nemotron-Personas-India: A dataset of synthetic human figures from India.
* Use online:
Nemotron-Personas-India is a synthetic human dataset for India released by NVIDIA in 2025. It aims to improve the diversity of synthetic data, reduce model bias, and prevent model collapse by reflecting the real geographical and demographic distribution of India. It is mainly used to support Indian model developers in building sovereign AI systems that can incorporate regionally specific demographic features and cultural backgrounds.
This dataset contains 3 million records, each with 7 virtual character roles, totaling approximately 21 million character descriptions and about 7.7 billion tokens, of which 2.9 billion tokens are character-related information. The data covers 36 states and union territories of India, 640 districts, and includes approximately 560,000 unique names. It is provided in three language versions: English, Devanagari, and Latin, with approximately 1 million records in each language version. This dataset is based on real demographic and geographical distribution information from the 2011 Indian census and election registers. All characters are completely virtual and do not correspond to any real personal identity information; data on minors under the age of 18 has also been excluded.
12. Nemotron-Personas-Belgium (Belgian synthetic human dataset)
* Use online:
Nemotron-Personas-Belgium is a dataset of synthetic people from Belgium released in 2026 by NVIDIA in collaboration with Pleias and KU Leuven. It aims to comprehensively reflect the diversity and characteristics of the Belgian population, primarily to expand the diversity of sovereign AI synthetic data, mitigate data and model bias, and improve the diversity of model responses.
This dataset contains 1.2 million records, each with 6 virtual character roles, totaling approximately 1.8 million character descriptions and about 1.9 billion tokens, of which 867 million tokens are character-related information. The data covers 581 Belgian municipalities and 3 administrative districts, containing approximately 260,000 unique names. It is available in four languages: Dutch, French, German, and English, with 300,000 records in each language version. This dataset is based on Belgian 2021 census data and 2025 population structure data. All characters are completely virtual and do not correspond to any real personal identity information; data on minors under the age of 18 has also been excluded.
13. Nemotron-Personas-Vietnam: A dataset of synthesized Vietnamese people.
* Use online:
Nemotron-Personas-Vietnam is a dataset of synthetic Vietnamese people released by NVIDIA in 2026. It aims to comprehensively reflect the diversity and characteristics of the Vietnamese population and is mainly used to support the development of AI models for Vietnamese sovereignty, mitigate data bias, and improve the diversity of model responses in the Vietnamese cultural context.
This dataset contains 100,000 records, each with 6 virtual character roles, totaling approximately 600,000 character descriptions and about 118 million tokens, of which 52 million tokens are character-related information. The data covers 6 centrally administered municipalities and provinces in Vietnam and includes approximately 13,000 unique names. This dataset was generated based on official Vietnamese statistics and localized expert knowledge provided by the FPT Group. All characters are completely virtual and do not correspond to any real personal identity information; data on minors under the age of 18 has also been excluded.
14. Nemotron-Personas El Salvador (Salvadoran synthetic human dataset)
* Use online:
The Nemotron-Personas-El Salvador dataset, released by NVIDIA in 2026, is a dataset of synthetic El Salvadoran people. It aims to comprehensively reflect the diversity and characteristics of the El Salvadoran population and is primarily used to support the development of AI models for El Salvadoran sovereignty, mitigate data bias, and enhance the diversity of model responses within the El Salvadoran cultural context.
This dataset contains 148,000 records, totaling approximately 1 million character descriptions and about 300 million tokens, of which 161 million tokens are character-related information. The data covers 14 provinces and 44 cities in El Salvador and includes approximately 144,000 unique names. This dataset is based on real demographic, geographical distribution, and occupational statistics from El Salvador's 2024 Seventh National Population Census and Sixth National Housing Census. All characters are completely virtual and do not correspond to any real personal identity information; data on minors under the age of 18 has also been excluded.
15. Nemotron-Personas Singapore (Singapore Synthetic Person Dataset)
* Use online:
Nemotron-Personas-Singapore is a dataset of synthetic people from Singapore released by NVIDIA in 2026. It aims to reflect the real-world population, geographical, and personality characteristics of Singapore. Its primary purpose is to support the development of sovereign AI systems, enhance the diversity of synthetic data, mitigate model bias, and prevent model collapse.
This dataset contains 148,000 records, each with 6 virtual character roles, totaling 888,000 character descriptions and approximately 118 million tokens, of which 48 million tokens are character-related information. The data covers 55 planning areas in Singapore and includes approximately 146,000 unique names. This dataset was generated based on the 2024 Singapore Census, name authorization data from the National Library Board (NLB), and official information provided by the Housing & Development Board (CEA). All characters are completely virtual and do not correspond to any real personal identity information; data on minors under the age of 18 has also been excluded.
