Machine-learning methods have enabled RNA modification detection from nanopore direct RNA sequencing. However, the existing nanopore-based RNA modification detection tools are limited, as each modification model requires a large amount of data and compute resources for training. Here we developed arowana, a transformer-based training framework for basecaller and RNA modification detection. We trained arowana modification callers and showed their ability to detect nine modifications stemming from the four nucleotide bases accurately. This demonstrates arowana’s potential to be expanded to other modifications.