Hm. Not much of an improvement over the method in your first post, but it's quite a significant change when compared to old method and old Lua version (earlier in the thread). Nice job.
So that is a 31% profit if I am correct. How many Lua files are there and how much overall performance profit could you have?
Quite impressive that you have made such progress, but if there are very few Lua files (I have no idea) to load, the overall loading time would not be significant effected.
The dataset tables have no reason for being lua, but replacing the flyff scanner scripts with lua is totally reasonable.
Sure, the datasets are read with the scanner but it's not to be considered as a script.