[lunixbochs]

A handful of the datasets I tested are fully held out (I have reason to believe none of the models have trained on them), and talon was trained on none of the dev or test data of any of the datasets in question.

Due to whisper's weakly supervised training on a large amount of automatically scraped data and reliance on a bigger language model, it's far more likely whisper had seen some of the test data before.