Upload train split with audio + prediction_text + reference_text (100 samples, 20–55s)
initial commit