A major benefit of Model Driven Engineering (MDE) relies on the automatic generation of artefacts from high-level models through intermediary levels using model transformations. In such a process, the input must be well-designed and the model transformations should be trustworthy. Due to the specificities of models and transformations, classical software test techniques have to be adapted. Among these techniques, mutation analysis has been ported and a set of mutation operators has been defined. However, mutation analysis currently requires a considerable manual work and suffers from the test data set improvement activity. This activity is seen by testers as a difficult and time-consuming job, and reduces the benefits of the mutation analysis. This paper addresses the test data set improvement activity. Model transformation traceability in conjunction with a model of mutation operators, and a dedicated algorithm allow to automatically or semi-automatically produce test models that detect new faults. The proposed approach is validated and illustrated in a case study written in Kermeta.