In software development, defects affect quality and cost in an adverse way. Therefore, various studies have been conducted defect prediction. Among them, prediction using another project is important but it is one of the most difficult problem. In this research, we conducted experiments that predict defects in Eclipse BIRT project using tokens from Eclipse and Netbeans projects as a training data by fault-prone filtering. Fault-prone filtering uses tokens in modules to classify the modules. Thus, investigating tokens are very important to improve prediction method. The result shows that fault-prone filtering can predict defects using another project tokens as a training data. In addition, we confirmed that the distribution of probability of faulty of tokens is significantly more important than the number of overlap between training tokens and test tokens.