I am working on improving an existing deep learning model which was published in last year A* conference. The code and datasets are made available on GitHub. When I ran them the accuracy was less than what they cited on paper I ran code 10 times answer changed a little(>1%) but the difference of cited score and actual scores are like( approx 5%). what should I report in my paper actual one or one I am getting   

P.S: I tried contacting all the authors and also opened an issue on github repository of code