Benchmarking Large Language Models with Augmented Instructions for Fine-grained Information Extraction