As the other commenter said, this is incorrect. The input was a sequence of legal moves (not even "real" moves - most of the training data was synthetically generated with "generate legal moves" as the only constraint).

Deducing board state from this is extremely non-trivial.