HiiragiLee/affectscore-ace-step-r64-20260629

AffectScore -- LoRA r=64 (diverged -- not recommended)

LoRA adapter for ACE-Step v1 3.5B from the AffectScore rank sweep (r ∈ {16, 32, 64}). This checkpoint diverged at epoch 40 following a gradient spike caused by Q2-weighted sampling concentrating the 34 available Q2 clips into consecutive batches. It is provided for reproducibility only. Use r=32 instead.

Code and Colab notebooks: github.com/LeeTgk/affectscore (link active after publication)

---

Training

Parameter	Value
Base model	ACE-Step/ACE-Step-v1-3.5B
LoRA rank	64
LoRA targets	cross_attn.q_proj, cross_attn.v_proj
use_rslora	True
lora_dropout	0.05
Training clips	1,571 (CC0/CC-BY, CLAP-quality filtered)
Epochs	50 (diverged at epoch 40)
Learning rate	1e-4
Batch size	8
Hardware	NVIDIA A100 (Google Colab)

Training dataset: https://zenodo.org/records/21830658

---

Ablation variants

See affectscore-ace-step-r32-20260629 for full model details, evaluation results, and the complete variant table.

---

Citation

Citation will be added once the paper is published. If you use this model before then, please link to this repository.