Towards Faithful and Robust LLM Specialists for Evidence-Based Question-Answering