NotaGen explores using Large Language Model training techniques for symbolic music generation, specifically producing high-quality classical sheet music. The model uses a three-stage approach: pretraining on 1.6M musical pieces, fine-tuning on ~9K classical compositions with period-composer-instrumentation prompts, and reinforcement learning via the novel CLaMP-DPO method that requires no human annotations or predefined rewards. The resulting model is published on HuggingFace with an interactive demo.