If you are getting vastly different results in error rates on different runs of fit_one_cycle while keeping the model, data and approach constant, is it a viable approach to just keep running it until you parameters with a suitably low error rate or is this an indication that we our model is not a good one for predicting on the data?
The different errors are due to the various data augmentation techniques used. Generally, it does not matter in most of the cases. Just train a bit longer.