Which corpora were used for training?

[Menschenkindlein]

Hi,
I'm trying to train a model with comparable performance. I'm using ontonotes 5 and questionbank, converted using clearnlp-3.1.2, but my model is significantly worse in many cases. It is especially clear on questions. As I understand it, it's because questionbank has simpler constituency tags (e.g., just NP, instead of NP-SBJ-2), which prevents high-quality conversion to dependency. Did you use manually annotated/reviewed corpora? Is it possible to achieve the same level of quality using only open corpora?
Thank you in advance.