It does text to Morse well, it explains its interpretations of Morse as text quite well, but consistently, even though it repeats the entered Morse sequence correctly, the series of dots and dashes it uses in the explanation don’t match what was entered. (I suppose its also possible it normalizes write and its training set has too little Morse-to-text so the input-output relationship is not captured in the model strongly enough.
Morse needs some concept of understanding time (Long vs short)
Human can differentiate the longs vs shorts and when a word is finished easily. I think you need like two gpt for that. One ist translating from Morse to NL which will be processed by the second gpt