Our understanding of the ways likes and dislikes are represented in memory is limited. The present work examines the information encoded during evaluative conditioning (EC). EC concerns the acquisition of attitudes via stimulus pairings. While EC is commonly viewed to result in links between stimulus identities (stimulus-stimulus learning), we argue that learners might abstract away from specific stimuli and rather store their overall valence (“positive”, “negative”; stimulus-valence learning). In particular, this might be the case when the overall US valence rather than the specific US can be predicted with high accuracy during EC. Two experiments (N = 522) modified an EC procedure in a way that predictions can be made and tested throughout learning. Every conditioned stimulus (CS) was either followed by multiple USs of the same valence (forward procedure), or multiple USs of the same valence were followed by the same CS (backward procedure). In the backward procedure, each US predicts one specific CS. In the forward procedure, each CS predicts the overall US valence, but cannot predict the specific US. A US revaluation procedure was implemented to test the format of acquired US representations. Results show that presentation sequence did not influence the degree of US revaluation. However, US revaluation was observable in the first experiment employing multiple repetitions, but not in the second one with single learning trials. We discuss implications of the findings and consider them in the light of reinforcement learning mechanisms.