Questions regarding Human SL model

[pnprog]

Thanks for releasing the "Human SL", I think there is a lot of potential to use that model, and I have been experimenting a lot with it recently.

I have some question regarding specifically the model file used (v1.15.0/b18c384nbt-humanv0.bin.gz):

• How is this file built compared to classic models? is it like a collection of models (one for each profile) stacked together? Is there a tool I can use to "inspect" it?
• What is the "block size" of this model? like 40-block? 60-block?
• How easy or hard would it be to expand the number of profiles that make that model? For instance, if I wanted to add the profiles preaz_30k to preaz_21k (or proyear_1700 to proyear_1799) to the existing list of models? (given that I already have an existing trained model, or the game collections to train it if necessary)

One last question, regarding the profiles made from older games:

• For profiles that have been trained on games record from before the current komi value, or even before the introduction of komi (for instance proyear_1900), can we expect the model to be at a disadvantage because of that? For instance, could the model believe it is loosing the game as white, because it "does not know" it has a 7.5pt komi?

Thanks a lot!