Novel contribution: LLM-based minority class augmentation via Groq (llama-3.3-70b)

with self-consistency filtering for label fidelity.

Test Accuracy: 0.9265 | Macro F1: 0.8926

Baseline (no augmentation): Acc=0.9235 | F1=0.8831

ICCA 2026 HCI Research — MoodShift Adaptive Chatbot