Skip to content

AI companies

How does Nvidia source training data?

By SourceX Editorial · Updated

Short answer

Nvidia publishes research describing how it builds training datasets, such as its Nemotron-CC work refining Common Crawl web data, and offers licensed and synthetic data tools. Nvidia also faces a class action lawsuit from authors alleging some of its earlier models were trained using copyrighted books obtained from shadow libraries. This page reflects Nvidia's own publications and the public litigation record.

Nemotron-CC and Common Crawl#

Nvidia researchers published a paper on Nemotron-CC, describing how they transformed Common Crawl, an existing public web-scrape dataset, into a refined pretraining dataset by combining classifier filtering, synthetic data rephrasing and reduced heuristic filtering.

Author copyright litigation#

A group of authors, including Abdi Nazemian, Brian Keene, Stewart O'Nan, Andre Dubus III and Susan Orlean, filed a class action complaint against Nvidia in the Northern District of California alleging that Nvidia trained Megatron-family models, including Megatron 345M and Nemotron-4 15B, using copyrighted books allegedly obtained from shadow libraries such as Anna's Archive.

A federal court order on Nvidia's motion to dismiss was granted in part and denied in part, meaning some of the plaintiffs' claims were allowed to proceed, according to the court's own order.

  1. 01Your work records
  2. 02Personal details removed
  3. 03You approve the release
  4. 04Licensed to AI developers
How it works with SourceX: Your work records → Personal details removed → You approve the release → Licensed to AI developers.Illustrative

What Nvidia has not disclosed#

Nvidia has not published a comprehensive breakdown of every dataset used across its commercial Nemotron model family; most public detail comes from individual research papers on specific dataset components like Nemotron-CC.

What this means for a business#

A business licensing infrastructure or models from Nvidia should note that Nvidia's role spans both training its own models and supplying tools and frameworks other companies use to build and train their own models, so data-sourcing questions may need to be asked separately for each use case.

SourceX and Nvidia#

SourceX is not affiliated with Nvidia. This page summarizes public reports; it does not mean Nvidia buys data through SourceX.

Frequently asked questions

Has Nvidia been sued over its training data?

Yes, authors filed a class action in the Northern District of California alleging Nvidia trained certain Megatron-family models on copyrighted books from shadow libraries; a court order in the case was granted in part and denied in part.

Does Nvidia use Common Crawl to build datasets?

Yes, Nvidia published research on Nemotron-CC describing how it refined Common Crawl data into a pretraining dataset.

Does Nvidia only train its own models?

No, Nvidia also builds tools and frameworks, such as NeMo Megatron, that other companies use to process and train on their own data.

Sources

Public sources, checked 2026-10-09. SourceX is not affiliated with the organizations named.

  1. Nemotron-CC paper describing Common Crawl refinement for pretraining arxiv.org
  2. Authors' class action complaint alleging Nvidia trained Megatron models on copyrighted books from shadow libraries storage.courtlistener.com
  3. Court order granting in part and denying in part Nvidia's motion to dismiss govinfo.gov

Related resources

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies Talk to SourceX
See if your company qualifies