Yeah, that happens a lot. The model starts fine with num2words, but in longer chats it falls back to digits like 1-0-0-0. Best fix is to re-run the number-to-words conversion (or a regex check) right before sending each reply so it stays consistent.