Modified version of fairseq, including new implementations for criterions using reinforcement learning methods. The implementation of two versions of RL criterions are in Version 1 and Version 2, respectively. Some experiments have been run with the new criterions. The results and conclusions are reported here.