HiiragiLee/affectscore-ace-step-r16-20260629

AffectScore -- LoRA r=16

LoRA adapter for ACE-Step v1 3.5B fine-tuned for affect-conditioned game music generation. Lower-capacity rank variant from the AffectScore rank sweep (r ∈ {16, 32, 64}). r=32 was selected by held-out MER accuracy; this checkpoint is provided for reproducibility of the rank sweep.

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	16
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
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.