fix: reduce number of languages, add lora

feat: switch train set -> better perf

fix: max len tokenizer

add tokenizer