I wondered if those error rates were proportional to Github's growth over time, so I looked it up. It seems that they have 40M users in 2019[1] and 73M users in 2021[2], which translates to 0.975 incidents per million users per year in 2019 compared to 1.178 in 2021.
So perhaps they are not exactly improving, but maybe there is some other way to normalize the data.