Our social communications and the expression of our beliefs and thoughts are becoming increasingly mediated and diffused by online social media. Beyond countless other advantages, this democratization and freedom of expression is also entailing the transfer of unpleasant offline behaviors to the online life, such as cyberbullying, sexting, hate speech and, in general, any behavior not suitable for the online community people belong to. To mitigate or even remove these threats from their platforms, most of the social media providers are implementing solutions for the automatic detection and filtering of such inappropriate contents. However, the data they use to train their tools are not publicly available.In this context, we release a dataset gathered from Mastodon, a distribute online social network which is formed by communities that impose the rules of publication, and which allows its users to mark their posts inappropriate if they perceived them not suitable for the community they belong to. The dataset consists of all the posts with public visibility published by users hosted on servers which support the English language. These data have been collected by implementing an ad-hoc tool for downloading the public timelines of the servers, namely instances, that form the Mastodon platform, along with the meta-data associated to them. The overall corpus contains over 5 million posts, spanning the entire life of Mastodon. We associate to each post a label indicating whether or not its content is inappropriate, as perceived by the user who wrote it. Moreover, we also provide the full description of each instance. Finally, we present some basic statistics about the production of inappropriate posts and the characteristics of their associated textual content.