IEPile is a large-scale, high-quality bilingual (Chinese and English) information extraction (IE) instruction fine-tuning dataset developed by Zhejiang University, covering three core subtasks: named entity recognition (NER), relation extraction (RE), and event extraction (EE). The dataset contains about 2 million instruction samples, totaling about 320 million tokens, covering multiple fields such as general, medical, and financial.

The research team carefully integrated 26 English and 7 Chinese IE datasets and adopted the proposed "schema-based polling instruction construction method", including the construction of a hard negative sample dictionary and polling instruction generation, to ensure the high quality of the dataset. The construction of IEPile significantly improved the performance of large models in information extraction tasks, especially zero-shot generalization capabilities, and provided valuable resources for information extraction research.