LLM vs classical ML across 8 datasets: labeled data still favors SVM and XGBoost

Ok_Juggernaut2187 · reddit · 2026-09-20

The author benchmarked Jev in zero-shot and few-shot settings against 11 classical ML models (SVM, XGBoost, logistic regression, etc.) on 8 classification datasets using balanced accuracy. Takeaways: with labeled data, classical models win — they beat Jev on Banking77 and business tabular data while avoiding per-call API costs; Jev only did well on IMDb; few-shot examples gave no consistent boost; and there's no reason to replace a working ML model. Setup: 3 training seeds, fixed test split, limited tuning budget, with zero-shot predictions cached across seeds. Full results, code and notebook are open-sourced.

Original post →

More from Models

Models channel →