The paper introduces Pattern-Exploiting Training (PET), a semi-supervised training method that combines task descriptions with pre-trained language models to improve few-shot learning performance. PET reformulates input examples into cloze-style phrases, which are then used to assign soft labels to a large set of unlabeled examples. Standard supervised training is performed on the resulting soft-labeled dataset. The method is evaluated on various tasks and languages, showing significant improvements over unsupervised and strong semi-supervised approaches, especially in low-resource settings. An iterative variant, iPET, further enhances performance by training multiple generations of models on increasingly larger datasets. The paper also discusses the effectiveness of auxiliary language modeling and the importance of knowledge distillation in improving overall accuracy.The paper introduces Pattern-Exploiting Training (PET), a semi-supervised training method that combines task descriptions with pre-trained language models to improve few-shot learning performance. PET reformulates input examples into cloze-style phrases, which are then used to assign soft labels to a large set of unlabeled examples. Standard supervised training is performed on the resulting soft-labeled dataset. The method is evaluated on various tasks and languages, showing significant improvements over unsupervised and strong semi-supervised approaches, especially in low-resource settings. An iterative variant, iPET, further enhances performance by training multiple generations of models on increasingly larger datasets. The paper also discusses the effectiveness of auxiliary language modeling and the importance of knowledge distillation in improving overall accuracy.