AI-assisted protein engineering means using machine learning models to mine mutants of a protein from its huge mutant library. Existed techniques, such as supervised models and zero-shot models, have received a lot of attention. The supervised approaches require a lot of data for optimal performance while the zero-shot models do not need assay-labeled data but their performance cannot be improved with labeled data increasing. In practice, researchers only have a small number of data, which is more like a few-shot learning problem. However, the few-shot learning within this domain has remained relatively unexplored. Moreover, there exist no standardized datasets and benchmarks that offer comprehensive basis for evaluating few-shot learning methods. In this paper, we present a few-shot learning dataset (FS-mutant) and benchmarking procedure. The dataset contains mutant data from eleven proteins with a variety of functions while the training set size are limited to 20, 40, 80, 160, and 320. In addition, we implement some zero-shot learning, supervised-learning, and meta-learning baselines and evaluate their performance on FS-mutant. We hope that FS-mutant will encourage future work on this extremely challenging domain. Data, code, and scripts can be found at https://github.com/ginnm/fs-mutant.