Data

Before a benchmark result means anything, two questions have to be settled: whether the test set leaked into training, and whether the sample was large enough to distinguish the systems being compared. Both are routinely skipped.