Are All the Datasets in Benchmark Necessary? A Pilot Study of Dataset Evaluation for Text Classification