Clustering-Based Synthetic Data Distillation Boosts Compact Finance Models

A distillation framework for financial sentiment analysis uses clustering-based seed selection to generate synthetic training data from small labeled sets, then trains compact student models from a large instruction-tuned teacher. On noisy financial text, the compact model trained on full synthetic-seed data matches or exceeds baselines with minimal real labeled examples.