WeReCooking/ACE-Step-xl-sft-LoRA-1374files

ACE-Step XL-SFT LoRA — 1374 files, 100 epochs ⚠️ UNTESTED (no working inference)

PEFT LoRA adapter trained on the xl-sft variant (4.99B AceStepConditionGenerationModel) using a curated 1374-file music dataset.

The LoRA itself loaded correctly via PEFT and trained without issues. However, no working inference path for xl-sft exists yet — both the base xl-sft model AND the base+LoRA produce repetitive output through the available generate_audio() function. The bug is in the inference code, not in this LoRA. Until xl-sft inference is fixed, the LoRA quality cannot be evaluated.

Two checkpoints included

Folder	Epoch	Best MA5	Notes
best-ep67/	67	0.8371	Best validation MA5 (recommended)
final-ep100/	100	0.8388	Last epoch (loss plateaued at ep60)

Training config

• Variant: xl-sft (4.99B params)
• LoRA: rank=64, alpha=128, dropout=0.1, target_modules=[q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj], target=both (self_attn + cross_attn) + MLP
• Optimizer: adamw8bit, lr=5e-5, batch=4 (effective 16 with grad_accum=4), warmup=300
• Schedule: cosine, 100 epochs, target-loss=0.5 (cruise control), save_every=10
• Data: 1374 audio files, random 60s crop per epoch

Loss curve

• ep1: 0.852 -> ep50: 0.840 -> ep67: 0.837 (best) -> ep100: 0.836
• Plateaued from epoch 60 onwards

Inference

To use locally with the generate_audio() function from WeReCooking/ACE-Step-CPU: set variant="xl-sft", adapter_path="path/to/best-ep67". Requires local GPU + xl-sft checkpoint — NOT compatible with the Space itself (turbo GGUF). The xl-sft inference path in that function currently produces repetitive output for both base and LoRA — needs investigation.