fix: reduce number of languages, add lora

feat: switch train set -> better perf

add tokenizer

chore: initial commit