Detoxify-IT: An Italian Parallel Dataset for Text Detoxification

AbstractToxic language online poses growing challenges for content moderation. Detoxification, which rewrites toxic content into neutral form, offers a promising alternative but remains underexplored beyond English. We present Detoxify-IT, the first Italian dataset for this task, featuring toxic comments and their human-written neutral rewrites. Our experiments show that even limited fine-tuning on Italian data leads to notable improvements in content preservation and fluency compared to both multilingual models and LLMs used in zero-shot settings, underlining the need for language-specific resources. This work enables detoxification research in Italian and supports broader efforts toward safer, more inclusive online communication.