Improving Model Factuality with Fine-grained Critique-based Evaluator