The expansion of social networks has accelerated the transmission of information and news at every communities. Over the past few years, the number of users, audiences and social networking publishers, are increased dramatically too. Among the massive amounts of information and news reported on these networks, we are faced with issues that have not been verified which is called “rumors”. Identifying rumors on social networks is carried out in the form of rumor detection approaches; the massive amount of these news and information force to use the machine learning techniques. The most important problem with auto-detection approaches is the lack of a database of rumors. For that matter, in this article, a collection of rumors published on the social network “telegrams” have been collected. These data are gathered from five Persian-language channels that have specially reviewed this issue. The collected data set contains 3283 messages with 2829 attachments, having a volume of over 1.6 gigabytes. This dataset can also be used for different purposes of natural language processing.