We trained these skip-transcoders using signum, over 1B tokens. Trained with input and output normalized.