I finally got it to work, by excluding transcription_text. I think maybe they are too large for the model to handle