Comment by PeterStuer

That was also my first thought. The discrepancy is just too large to be the mere result of a transformer model fed more chess data.