One important idea behind implementing language models is that the probability distributions are not precomputed. Instead, the model only stores the raw counts of n-gram occurrences and then computes the probabilities on demand. This makes smoothing possible.
ETS data set is removed. The ETS data set is proprietary and licensed to Columbia University for research and educational use only (as part of the Linguistic Data Consortium. This data set is extracted from https://catalog.ldc.upenn.edu/LDC2014T06.