Beyond Accuracy Optimization: Computer Vision Losses for Large Language Model Fine-Tuning F i n d i n g s - E M N L P