NVIDIA’s Open‑Source Nemotron Datasets: 10 T+ Tokens, 40 M Samples Across Math, Code, and Multilingual Dialogue

When model size is no longer the sole bottleneck, the quality, structure, and task alignment of training data become the primary determinants of large‑language‑model performance. NVIDIA’s Nemotron data series supplies more than 10 T tokens and 40 M post‑training samples, covering the full lifecycle from foundational pre‑training to agentic workflows.

Nemotron‑CC‑v2 (Pre‑training)

Release: 2025. Builds on Nemotron‑CC with eight new Common Crawl snapshots (2024‑2025), global deduplication, English filtering, and Qwen3‑30B‑A3B‑generated diverse QA translated into 15 languages. Cited in “NVIDIA Nemotron Nano 2: An Accurate and Efficient Hybrid Mamba‑Transformer Reasoning Model”. Online URL: https://go.hyper.ai/KbOSx

Nemotron‑Pretraining‑SFT‑v1 (Supervised Fine‑tuning)

Release: 2025. Synthetic dataset for instruction following, reasoning, code, and general QA. Includes high‑quality math and science material, graduate‑level academic texts, and complex multi‑choice questions with full solutions covering math, code, general, and logical reasoning. Cited in the same Nemotron Nano 2 paper. Online URL: https://go.hyper.ai/nF9Hl

Nemotron‑Pretraining‑Code‑v1 (Code)

Release: 2025. Curated from GitHub with multi‑stage deduplication, license enforcement, and heuristic quality checks. Contains code Q&A pairs in 11 programming languages, 175.1 B synthetic code tokens and ~747.4 B metadata tokens. Cited in the Nemotron Nano 2 paper. Online URL: https://go.hyper.ai/37WQG

Nemotron‑Pretraining‑Code‑v3 (Incremental Code)

Release: 2025 for the Nemotron‑3 series. Adds 1.463 B new source‑code files collected up to 2025‑09‑30. Intended to be combined with v1 and v2 for a complete code corpus. Online URL: https://go.hyper.ai/QcGUu

Nemotron‑CC‑Math (Mathematics Pre‑training)

Release: 2025. 133 B tokens built from Common Crawl via NVIDIA Lynx and a lightweight LLM pipeline, preserving equation and code formatting and converting all math to editable LaTeX. Benchmark‑validated for high‑quality math pre‑training. Cited in “Nemotron‑CC‑Math: A 133 Billion‑Token‑Scale High Quality Math Pretraining Dataset”. Online URL: https://go.hyper.ai/6jNCq

Nemotron‑Personas‑Korea

Release: 2026. 1 M records, each with 7 virtual personas (≈7 M personas, ~1.7 B tokens, 1 B persona‑related). Covers 17 Korean provinces, 252 districts, 209 167 unique names. Generated from Korean statistical services, Supreme Court, health insurance, rural research, and NAVER Cloud. All personas are synthetic and exclude individuals under 18. Online URL: https://go.hyper.ai/49T4h

Nemotron‑Personas‑France

Release: 2026 (jointly with Pleias). 6 M French persona instances across 1 M records, each record contains 22 fields (e.g., scholar, athlete, artist, foodie, traveler). Online URL: https://go.hyper.ai/gCJBP

Nemotron‑Personas‑Brazil

Release: 2026 (jointly with WideLabs). 1 M records, 6 personas each (≈6 M personas), 14 fields per record. Covers all 26 Brazilian states plus the Federal District. Online URL: https://go.hyper.ai/9MYxx

Nemotron‑Personas‑USA

Release: 2025. ~1 M records, 6 personas each (≈6 M personas, ~600 k persona fields, 16 context fields). Covers all 50 states, Puerto Rico, and the Virgin Islands, with 29 k ZIP codes, 15.2 k cities, ~97 k unique names, and >560 occupations. Online URL: https://go.hyper.ai/VkjC8

Nemotron‑Personas‑Japan

Release: 2025. 1 M records, 6 personas each (≈6 M personas, ~1.4 B tokens, 0.85 B persona). Covers all 47 prefectures, 1.5 k occupations, and ~950 k unique names. Excludes individuals under 18. Online URL: https://go.hyper.ai/3oCJ3

Nemotron‑Personas‑India

Release: 2025. 3 M records, 7 personas each (≈21 M personas, ~7.7 B tokens, 2.9 B persona). Covers 36 states and union territories, 640 districts, ~560 k unique names, and three language versions (English, Devanagari, Latin). Online URL: https://go.hyper.ai/lT28T

Nemotron‑Personas‑Belgium

Release: 2026 (jointly with Pleias & KU Leuven). 1.2 M records, 6 personas each (≈1.8 M personas, ~1.9 B tokens, 0.867 B persona). Covers 581 municipalities, 3 regions, and four language versions (Dutch, French, German, English). Online URL: https://go.hyper.ai/epZ5X

Nemotron‑Personas‑Vietnam

Release: 2026. 100 k records, 6 personas each (≈600 k personas, ~118 M tokens, 52 M persona). Covers central cities and provinces, ~13 k unique names. Online URL: https://go.hyper.ai/g7Msj

Nemotron‑Personas‑El Salvador

Release: 2026. 148 k records, ~1 M personas, ~300 M tokens (161 M persona). Covers 14 provinces and 44 cities, ~144 k unique names. Online URL: https://go.hyper.ai/39ALO

Nemotron‑Personas‑Singapore

Release: 2026. 148 k records, 6 personas each (≈888 k personas, ~118 M tokens, 48 M persona). Covers 55 planning areas and ~146 k unique names. Online URL: https://go.hyper.ai/84YYI