Model Lab — severity
Severity-classification benchmark: LLM (DeepSeek) vs. classic ML models vs. a fine-tuned German BERT — all on the same 150 test cases.
Try it live
Runs entirely in your browser (exported Logistic Regression). Type a German note.Benchmark
On synthetic data the trained models score high because TF-IDF hands them the generator's phrase library (memorization) — the 17-point gap between LightGBM and XGBoost, two near-identical methods, is the tell. None beats the 88.7% ceiling for these 150 cases; DeepSeek generalizes to real notes.