GPU-accelerated data curation for training better AI models, faster. Scale from laptop to multi-node clusters with modular pipelines for text, images, video, and audio.

Part of the NVIDIA NeMo software suite for managing the AI agent lifecycle.

• 2026-04: NeMo Curator 26.04 released with Cosmos-Xenna 0.2.0 upgrade, simplified Resources API, and Ray 2.54. See the release notes.
• 2026-02: NeMo Curator 26.02 released with Ray-based pipeline architecture for all modalities — text, image, video, and audio.

# Install for your modality
uv pip install "nemo-curator[text_cuda12]"

# Run the quickstart example
python tutorials/quickstart.py

Full setup: Installation Guide • Docker • Tutorials

NeMo Curator uses a modular, Ray-based pipeline architecture. Data flows through composable processing stages — each stage handles a discrete curation task (loading, filtering, deduplication, etc.) and can be configured with independent resource requirements.

Process and curate high-quality text datasets for large language model (LLM) training with multilingual support.

Curate large-scale image datasets for vision language models (VLMs) and generative AI training.

Process large-scale video corpora with distributed, GPU-accelerated pipelines for world foundation models (WFMs).

Prepare high-quality speech datasets for automatic speech recognition (ASR) and multimodal AI training.

High-quality training data is the single most important factor in building performant AI models. Raw datasets contain noise, duplicates, low-quality content, and potentially harmful material that degrade model performance and increase training costs.

At scale, data curation is a throughput maximization problem. A typical pipeline chains stages with very different compute profiles — lightweight CPU tokenization, small GPU classifiers, large GPU inference models — and a naive sequential approach leaves most hardware idle most of the time.

Example: Consider a pipeline with language identification (0.5B model, 1 GB VRAM, 2s/sample), tokenization (CPU-only, 1s/sample), and a 5B answer model (10 GB VRAM, 10s/sample) processing 1,000 questions on a single 102 GB GPU:

NeMo Curator achieves this by streaming data through the pipeline so all stages run concurrently, auto-balancing replicas to match each stage's throughput (2× language ID, 1× tokenizer, 10× answer model), and keeping GPU workers busy over 99% of the time after an initial warm-up period. See the scaling concepts for details.

NeMo Curator powers the data pipelines behind NVIDIA Nemotron models. For example, the Nemotron-4 pre-training dataset was curated using NeMo Curator's text processing pipeline across 8+ trillion tokens of multilingual web data, applying quality filtering, deduplication, and domain classification at scale.

NeMo Curator leverages NVIDIA RAPIDS™ libraries such as cuDF, cuML, and cuGraph along with Ray to scale workloads across multi-node, multi-GPU environments.

Proven Results:

• 16× faster fuzzy deduplication on 8 TB RedPajama v2 (1.78 trillion tokens)
• 40% lower total cost of ownership (TCO) compared to CPU-based alternatives
• Near-linear scaling from one to four H100 80 GB nodes (2.05 hrs → 0.50 hrs)

Real-World Recipe: The Nemotron-CC curation pipeline uses NeMo Curator end-to-end — from Common Crawl extraction through language identification, exact/fuzzy/substring deduplication, ensemble quality classification, and LLM-based synthetic data generation — to reproduce the Nemotron-CC datasets. The SDG stage is also available as an in-repo tutorial.

Data curation modules measurably improve model performance. In ablation studies using a 357M-parameter GPT model trained on curated Common Crawl data:

Results: Progressive improvements in zero-shot downstream task performance through text cleaning, deduplication, and quality filtering stages.

We welcome community contributions! Please refer to CONTRIBUTING.md for guidelines.

If you find NeMo Curator useful in your research, please cite:

@misc{nemo_curator,
  title = {NeMo Curator: GPU-Accelerated Data Curation for Training AI Models},
  author = {NVIDIA},
  year = {2024},
  url = {https://github.com/NVIDIA-NeMo/Curator}
}

For the data curation pipeline behind Nemotron models, please also cite:

@article{parmar2024nemotron4,
  title = {Nemotron-4 15B Technical Report},
  author = {Parmar, Jupinder and Satheesh, Shrimai and others},
  journal = {arXiv preprint arXiv:2402.16819},
  year = {2024}
}