Unimol Docking V2 strange result on posebuster set

I followed the posebuster.ipynb provided in your repo, downloaded the weight file and downloaded the eval_set zip files. However, since the eval_set does not contain pdb file, so I downloaded from the Posebuster repo and move the pdb file into the eval_set folder.
I ran everything smoothly, and the notebook end up giving me a RMSD<2A passing rate of something near 0.8.

However, when I export the predicted pose, and run the Posebuster quality check again, the passing rate significantly drops to lower than 0.6, and the passing rate with PB-valid drop below 0.2. I checked some of the prediction, in some part of the molecule, atoms clash together, which is quite surprising.

I wonder if there is any difference in the passing criteria in the Posebuster repo and the passing criteria written in your ipynb? And if there is details I missed in running the ipynb provided in your github?