YuE2-3B classical, v2 (full AR fine-tune + classical decoder LoRA), personal test
checkpoints/latest/ar_weights.safetensors and snapshots/step-*.safetensors hold the fine-tuned AR layer weights (bf16, names as in YuE2ForCausalLM). decoder/ holds the classical decoder (NAR) LoRA in nar_lora_joint_v9 format. Made with the molab notebook yue2_v2_molab.py. Non-commercial (YuE2 weights are CC BY-NC 4.0; some training audio is NC/ND).
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for orangedino2/yue2-classical-v2
Base model
m-a-p/YuE2-3B