Label Factory - LLMs for Training Small Language Classification Models at Scale

Learn how to build scalable content moderation systems using large language models to train small classification models through this 17-minute conference talk from Weights & Biases. Discover Zefr's patent-pending Label Factory process that creates small, multimodal, multilingual classifiers for Fortune 100 brands and advertisers without requiring human-annotated labels first. Explore the knowledge-distillation approach that generates unsupervised classification labels achieving over 90% accuracy compared to human counterparts while enabling collection at scale. Understand how social media content moderation faces significant challenges beyond inference scaling, including the need to continuously create new models as underlying policies change over time. Examine the integration of in-house infrastructure with Weights & Biases to streamline the creation, evaluation, and deployment of small classifiers with dramatically reduced time and effort while maintaining quality standards.