CRAFT: Clustering Rubrics to Diagnose Weak LLM Capabilities and Generate Targeted Fine-Tuning Data

Explain Like I'm Normal

CRAFT is a new framework that turns standard evaluation rubrics into a diagnostic map of model failures. Instead of just giving a score, it clusters grading criteria into a hierarchical tree to pinpoint exactly which capabilities are lacking and generates the specific data needed to fix them. This shifts model improvement from random vibes to targeted engineering.