Training Large Language Models (LLMs) requires processing massive-scale datasets efficiently. Traditional CPU-based data pipelines struggle to keep up with the exponential growth of data, leading to bottlenecks in model training. In this talk, we present NeMo Curator, an accelerated, scalable Python-based framework designed to curate high-quality datasets for LLMs efficiently. Leveraging GPU-accelerated processing with RAPIDS, NeMo Curator provides modular pipelines for synthetic data generation, deduplication, filtering, classification, and PII redaction—improving data quality and training efficiency.

We will showcase real-world examples demonstrating how multi-node, multi-GPU processing scales dataset preparation to 100+ TB of data, achieving up to 7% improvement in LLM downstream tasks. Attendees will gain insights into configurable pipelines that enhance training workflows, with a focus on reproducibility, scalability, and open-source integration within Python's scientific computing ecosystem.