Using a Local LLM as a Zero-Shot Classifier

A locally hosted Large Language Model can serve as an effective zero-shot classifier for messy, unstructured text data. This approach overcomes the limitations of traditional clustering methods by understanding semantic meaning rather than just surface-level features like keywords. The proposed pipeline involves defining candidate categories, prompting a local model like Gemma 2 to classify each text entry, and analyzing the resulting distribution. This technique was successfully used to categorize thousands of security annotations without labeled data, revealing key themes that were previously hidden in the free-text. While not suitable for all scenarios, it is a powerful tool for medium-scale classification tasks where domain knowledge is available.