PEGASUS is a novel pre-training objective for abstractive text summarization, designed to generate summary-like text from input documents. The method involves masking important sentences and generating these gaps from the remaining sentences, similar to extractive summarization. The pre-training objective, called Gap Sentences Generation (GSG), selects and masks whole sentences based on their importance to the document. The model is pre-trained on large text corpora, such as C4 and HugeNews, and evaluated on 12 downstream summarization tasks across various domains, including news, science, stories, instructions, emails, patents, and legislative bills. PEGASUS achieves state-of-the-art performance on all 12 datasets, measured by ROUGE scores, and shows superior performance in low-resource summarization tasks with only 1000 examples. Human evaluation further validates the quality of PEGASUS summaries, demonstrating that they achieve human-level performance on multiple datasets. The paper also explores the effects of pre-training corpora, gap-sentences ratios, vocabulary sizes, and hyperparameters, providing insights into the optimal configuration for abstractive summarization.PEGASUS is a novel pre-training objective for abstractive text summarization, designed to generate summary-like text from input documents. The method involves masking important sentences and generating these gaps from the remaining sentences, similar to extractive summarization. The pre-training objective, called Gap Sentences Generation (GSG), selects and masks whole sentences based on their importance to the document. The model is pre-trained on large text corpora, such as C4 and HugeNews, and evaluated on 12 downstream summarization tasks across various domains, including news, science, stories, instructions, emails, patents, and legislative bills. PEGASUS achieves state-of-the-art performance on all 12 datasets, measured by ROUGE scores, and shows superior performance in low-resource summarization tasks with only 1000 examples. Human evaluation further validates the quality of PEGASUS summaries, demonstrating that they achieve human-level performance on multiple datasets. The paper also explores the effects of pre-training corpora, gap-sentences ratios, vocabulary sizes, and hyperparameters, providing insights into the optimal configuration for abstractive summarization.