Google Research



One million English sentences, each split into two sentences that together preserve the original meaning, extracted from Wikipedia edits.

The WikiSplit dataset was constructed automatically from the publicly available Wikipedia revision history. Although the dataset contains some inherent noise, it can serve as valuable training data for models that split or merge sentences.