Russian might also be a BPE issue to some extent, but the flip side: the problems with using character-level encoding with narrow context window & small data: https://twitter.com/NineOfNein/status/1286738449660284928 (As long as you have a narrow context window, you’re stuck between a rock and a hard place.)
Russian might also be a BPE issue to some extent, but the flip side: the problems with using character-level encoding with narrow context window & small data: https://twitter.com/NineOfNein/status/1286738449660284928 (As long as you have a narrow context window, you’re stuck between a rock and a hard place.)