Stance detection is a classification task that determines whether a text is in favour, against or neutral towards a particular target. Arabic stance detection remains under-explored. This paper describes our work, which consists in evaluating a new dataset composed of user-generated texts in dialectal and formal Arabic on three targets: general labour union, illegal immigration, and administrative capital. We carried out experiments employing AraBERT-Twitter and Qarib transformers in addition to several machine learning classification models trained using different settings including target-specific and dialect-specific. The results show that training a model for each target, using Qarib, yields the best performance.