Configuration Parsing Warning:In adapter_config.json: "peft.task_type" must be a string

SoundSeed Multi-LoRA v1

Universal multi-genre LoRA adapter for ACE-Step v1.5 SFT music generation model.

Training Details

  • Base model: ACE-Step v1.5 SFT
  • LoRA rank: 128, alpha=256
  • Target modules: q_proj, k_proj, v_proj, o_proj, gate_up_proj, down_proj
  • Trainable params: 113.2M
  • Dataset: 3,323 multi-genre tracks (3,157 train / 166 val)
  • Learning rate: 5e-05 (cosine schedule with warmup)
  • CFG dropout: 10%
  • Sampling: WeightedRandomSampler for rare genres
  • Best epoch: 2 (early stopped at epoch 5, patience=3)
  • Best val_loss: 0.000667
  • Training time: ~5 hours on A100 80GB

Training Curve

Epoch Train Loss Val Loss
1 0.002058 0.000783
2 0.000799 0.000667
3 0.000786 0.000726
4 0.000783 0.000716
5 0.000698 0.000694

Framework

  • PEFT 0.18.1
Downloads last month
486
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support