CrowdStrike’s Journey in Customizing NVIDIA Nemotron Models for Peak Accuracy and Performance

CrowdStrike details its collaboration with NVIDIA to operationalize Nemotron models for security workflows, focusing on a natural language-to-CQL translator within the Falcon platform. The post outlines a pipeline including AST-based deduplication, high-accuracy PII scrubbing, synthetic natural language generation via NVIDIA NeMo Data Designer, and LoRA fine-tuning of Llama-3.3-Nemotron-Super-49B-v1.5, achieving 96% valid query accuracy and outperforming other leading LLMs. It also previews evaluation of the Nemotron 3 model family (Nano, Super, Ultra) for varied SOC use cases and explores safety and content moderation models to strengthen AI-driven security operations.