Nemotron-CC

license

NVIDIA-Open-Data

gated

manual(model-training-use)

dataset_card

present

Manually gated behind the NVIDIA Data Agreement for Model Training; synthetic portions may inherit generator-model license terms.

• https://huggingface.co/api/datasets/nvidia/Nemotron-CC-v2 recorded 2026-10-08

  Hub metadata: "gated":"manual" with the gate field "I agree to use this dataset for model training purposes ONLY".

• https://huggingface.co/datasets/nvidia/Nemotron-CC-v2 recorded 2026-10-08

  Dataset page gated on the agreement: "I confirm I have read and agree to NVIDIA Data Agreement for Model Training". The card's License/Terms of Use section names the "NVIDIA Data Agreement for Model Training" and says "The data may be freely used to train and evaluate with user agreement to open data license."

Adoption is measured by Hugging Face downloads of nvidia/Nemotron-CC-v2, the only usage channel on record for Nemotron-CC.

• https://huggingface.co/api/datasets/nvidia/Nemotron-CC-v2 recorded 2026-08-13

  45,663 downloads in the trailing 30 days for nvidia/Nemotron-CC-v2

An 8B model trained for 1T tokens on the high-quality subset scores 5.6 MMLU points above one trained on DCLM. The full dataset only matches DCLM on MMLU but holds about four times as many unique real tokens, and an 8B model trained for 15T tokens, 7.2T of them from it, beats Llama 3.1 8B by 5 MMLU points. Marin 8B moved its training mixture to Nemotron-CC in a later phase.

• https://arxiv.org/abs/2412.02595 recorded 2026-09-29

  arXiv abstract says an 8B model "trained for 15T tokens, of which 7.2T came from our dataset, is better than the Llama 3.1 8B model: +5 on MMLU", and that "using a high-quality subset of our data improves MMLU by 5.6 over DCLM" at 1T tokens while "our full 6.3T token dataset matches DCLM on MMLU".

• https://marin.readthedocs.io/en/latest/reports/marin-8b-retro/ recorded 2026-09-29

  Marin 8B retrospective says "Because we were running out of DCLM, we transitioned the data from our phase 1 and 2 mixture (DCLM+StarCoder+ProofPile 2) to a new mixture that was" Nemotron-CC and StarCoder.