Domain-Specific Fine-Tuning Outperforms GPT-4 in Korean Legal Classification

Fine-tuning the small-scale KLUE-BERT model on legal data achieved 99.3% accuracy in classifying Korean sexual offense precedents, outperforming GPT-3.5 and GPT-4.0. The results demonstrate that domain adaptation is more critical than raw parameter count for specialized legal text classification.