Overview of NeMo Curator

NeMo Curator is an open-source, enterprise-grade platform for scalable, privacy-aware data curation across text, image, video, and audio modalities.

NeMo Curator, part of the NVIDIA NeMo software suite for managing the AI agent lifecycle, helps you prepare high-quality, compliant datasets for large language model (LLM) and generative artificial intelligence (AI) training. Whether you work in the cloud, on-premises, or in a hybrid environment, NeMo Curator supports your workflow.

Target Users

• Data scientists and machine learning engineers: Build and curate datasets for LLMs, generative models, and multimodal AI.

• Cluster administrators and DevOps professionals: Deploy and scale curation pipelines.

• Researchers: Experiment with new data curation techniques and ablation studies.

• Enterprises: Ensure data privacy, compliance, and quality for production AI workflows.

How It Works

NeMo Curator speeds up data curation by using modern hardware and distributed computing frameworks. You can process data efficiently—from a single laptop to a multi-node GPU cluster. With modular pipelines, advanced filtering, and easy integration with machine learning operations (MLOps) tools, leading organizations trust NeMo Curator.

• Text Curation: Uses a pipeline-based architecture with modular processing stages running on Ray. Data flows through data download, extraction, language detection, rule-based quality filtering, deduplication (exact, fuzzy and semantic) and model based quality filtering.
• Image Curation: Uses pipeline-based architecture with modular stages for loading, embedding generation, classification (aesthetic, NSFW), filtering, and export workflows. Supports distributed processing with optional GPU acceleration.
• Video Curation: Employs Ray-based pipelines to split long videos into clips using fixed stride or scene-change detection, with optional encoding, filtering, embedding generation, and deduplication for large-scale video processing.
• Audio Curation: Provides ASR inference using models, quality assessment through Word Error Rate (WER) calculation, duration analysis, and integration with text curation workflows for speech data processing.

Key Technologies

• Graphics Processing Units (GPUs): Speed up data processing for large-scale workloads.
• Distributed Computing: Supports frameworks like Dask, RAPIDS, and Ray for scalable, parallel processing.
• Modular Pipelines: Build, customize, and scale curation workflows to fit your needs.
• MLOps Integration: Seamlessly connects with modern MLOps environments for production-ready workflows.

Concepts

Explore the foundational concepts and terminology used across NeMo Curator.