Good for text classification
gzip with a k-nearest-neighbor classifier. The method achieves competitive results with non-pretrained deep learning methods on six in-distribution datasets and outperforms BERT on all five out-of-distribution datasets, including four low-resource languages
