Unable to replicate results on BOOLEAN SATISFIABILITY PROBLEM

Thanks for sharing the code. 

I am not able to replicate the results on BOOLEAN SATISFIABILITY PROBLEM reported in Figure 4 of the paper. I use the provided bash script ```scripts/3-sat/train-mdm.sh``` with the 3sat5 dataset. After tuning the learning rate, the max accuracy I get is ~47%. However much higher number is reported in the paper. Can you please help me with this? 

Thank you.