🏥 Clinical LLM — QLoRA Fine-Tune on MedQA

Base model: Llama 3.2 3B Instruct  |  Fine-tune: QLoRA r=32  |  Dataset: MedQA (2,500 samples)  |  Accuracy: Base 10% → Fine-tuned 59.2%

These 6 questions were hand-selected from the validation set to give an honest picture of what fine-tuning achieves. The fine-tuned model is ~6x more accurate than the base model (59.2% vs 10%), largely because it learns to respond in the correct concise format rather than generating verbose explanations. That said, 59.2% accuracy means the model still gets a significant portion of questions wrong — the samples below reflect both the strengths and the honest limitations of a 3B parameter model fine-tuned on 2,500 examples. Try a few and see for yourself.

📊 Ablation Results

⚡ Quantization Benchmark (r=32, L4 GPU)