michelepapucci/impacts

IMPaCTS (Italian Multi-level Parallel Corpus for Controlled Text Simplification) is a research dataset for Italian text simplification. Text simplification is the task of rewriting complex sentences into plainer language while preserving meaning, useful for making public documents more accessible to general readers. This dataset provides training and evaluation data for building AI models that can do this for Italian. The dataset contains 1,066,828 sentence pairs. Each pair consists of a human-written original sentence from Wikipedia and Public Administration texts, alongside one or more machine-generated simplified versions. The average number of simplifications per original sentence is 9.6. The simplified sentences were generated automatically using an Italian language model prompted to produce multiple simplifications per input. Every row is annotated with readability scores for both the original sentence and its simplification. There are four readability scores per sentence: raw textual features (such as average characters per word), lexical features (vocabulary diversity), syntactic features (sentence tree depth and use of subordinate clauses), and an overall combined score. These scores come from the Read-it readability tool. The full set of extracted linguistic features is also included for both the original text and the simplification. The dataset is intended for researchers training or evaluating language models on controllable text simplification tasks. It is available as a zip archive in this repository (requires git-lfs to download) and also on HuggingFace. The repository accompanies a paper presented at LREC 2026.