Comment by argee

That’s the problem. If all (or zero) models were trained on it, it would be fine as a benchmark.