NVIDIA-NeMo/Curator

Scalable data pre processing and curation toolkit for LLMs

What it solves

NeMo Curator addresses the challenge of preparing massive datasets for AI training. It replaces ad-hoc scripts and notebooks with repeatable, GPU-accelerated pipelines that can scale from a single laptop to multi-node clusters, ensuring high-quality data through filtering, deduplication, and transformation.

How it works

The system uses a modular architecture based on stages (discrete tasks like loading, filtering, or classifying) and executors (the runtime environment).

  • Pipeline Composition: Stages are chained together, and the executor automatically scales replicas to maintain throughput.
  • Streaming Execution: It overlaps CPU and GPU workloads to keep GPUs highly utilized (often >99%).
  • Modality Support: It provides specialized building blocks for text, image, video, and audio data.
  • Distributed Scaling: It leverages NVIDIA RAPIDS (cuDF, cuML, cuGraph) and Ray to distribute workloads across multiple GPUs and nodes.

Who it’s for

ML engineers and data teams who need to process trillion-token scale datasets or multimodal data (images, video, audio) for training frontier AI models.

Highlights

  • Multimodal Capabilities: Dedicated pipelines for text (deduplication, language detection), images (aesthetic/NSFW filtering), video (scene detection, motion filtering), and audio (ASR transcription, WER filtering).
  • GPU Acceleration: Significant speedups over CPU-based alternatives, such as reducing fuzzy deduplication time from 10.7 hours to 0.65 hours.
  • Production Proven: Powers the data pipelines for NVIDIA Nemotron models, including the curation of 8+ trillion tokens of web data.
  • Synthetic Data Generation: Includes an OpenAI-compatible inference server to facilitate SDG and classification workflows within the pipeline.