Performance on factual question-answering benchmarks such as MedQA and GPQA Diamond can be restored to within one point of baseline after model sparsification and fine-tuning