This argument always come for every benchmark. But as I understand, training requires lots of examples. How will a lab go about getting that many very specific high quality examples of said benchmark?