I finetuned a LoRa on Qwen 3 30B A3B on data from my Claude export to see if I could make Qwen get rid of its frustrating, emoji-laden style.