Emir U • 3/30/2025

Bootstrapping ranking models with an LLM judge

This technical article details an experiment to bootstrap a personalized ranking model for Hacker News article titles. The author uses a Large Language Model (LLM) to label 500 titles based on a complex, user-written preference description. These labels are then used with sentence transformer embeddings and Ridge regression to create a predictive ranking model, achieving a 0.74 Spearman correlation with the LLM's judgments.

0 comments

#Machine Learning #llm #Sentence Transformers