We have proposed a new fault-prone prediction ap- proach, named “Fault-Prone Filtering [1],” using spam fil- tering technique. In the proposed approach, we predict fault-proneness by applying spam filter to raw source code modules. As for the training procedure of spam filtering, Training Only Errors(TOE) approach is widely used. In TOE, e-mail messages are classified in arrival order, and if misclassification happens with a message, it is trained to the corpus. We have reported that this procedure also works fine in the fault-prone filtering [1]. However, we perceived a problem that the accuracy of prediction is unstable and low at an early stage of TOE. It is necessary to keep high accuracy in order to adopt our model to software development. We thus try to reuse pre- trainied courpuses of previous similar project for a new project to improve the accuracy at an early stage of TOE.