Generative SLM Ensemble for NBME Clinical Span Extraction (0.732 F1). Novel 3-phase pipeline: (1) RAG-guided Qwen3-8B pseudo-labeling, (2) Sequential QLoRA fine-tuning of 4 diverse SLMs, (3) FSM-constrained inference + char-level voting. Outperforms baselines by +0.222 F1. Includes adapters, vLLM optimization for 2xT4, and full ablation code.
-
Updated
May 16, 2026 - Jupyter Notebook