First question, yes, the loading is real so that it can calibrate and calculate probability of every repeated word. But it's fast, none the less. This is a model Bigram LLM, that has learned from Terms and Conditions samplings. After it's loaded, type a word (preferably related to Terms and Conditions) in the field and hit enter. Then the LLM will predict the next few words, via the Markov Prediction (the Markov assumption says that the probability of a future word depends only on the current word), so this basic model is on the right track to be able to operate using bigram predictions using its corpus. Did any of that make sense? Anyway, here are some fun stats: As of 27/03/24, made with only 5 scripts, 132 blocks and a corpus of over 800 words. Prediction length goes up to 300 words, and the corpus is made from copy and pasted words from T&C's. Known Bugs: If the data has fewer spaces than the 'prediction length' variable, occasionally it results in a malfunctioned return :(
