Statistics of a Large-Scale Romanian Corpus for Language Modelling

We report here a series of detailed statistical analyses on a novel largescale multi-domain Romanian corpus that we use to train a small-language model. We identify the core vocabularies pertaining to six different domain-specific subcorpora and show that they follow the so-called Zipf’s law independent on how we count words. Moreover, we introduce two novel frequency-word maps for a domain-specific subcorpus, one showcasing word ranks and one measuring the deviation of the word structure from a perfect vowel-consonant or consonant-vowel repeating pattern. Finally, we show a few examples of prompt/response instances.