Compare against a benchmark corpus

[delcenjo]

Beyond train/val/test, it is useful to check a training set against a public benchmark to detect contamination. Accepting a benchmark file as a dedicated source and reporting contamination separately would support this.