LoRA fine-tune of google/gemma-2-2b-it for retrieval-augmented QA: answer from provided context, ignore distractor passages, and refuse when the context lacks the answer. Trained on general-knowledge (fineweb-edu) question/answer pairs with golden+distractor contexts.

Held-out eval (Gemini-Flash judge, /10): base closed-book 4.42, base+RAG 4.75, this model+RAG 6.30 (+1.88 vs base, +1.55 vs base+RAG, both p<0.001). Use it WITH retrieved context in the prompt.