Downloads · 30 days
9
12% of all-time downloads
affectscore-ace-step-r32-20260629-no-affect
HiiragiLee/affectscore-ace-step-r32-20260629-no-affect
affectscore-ace-step-r32-20260629-no-affect is a machine learning model from HiiragiLee. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as mit.
Ablation checkpoint for the AffectScore rank sweep. Identical to the full r=32 model except Layer 1 V-A conditioning is removed: every training clip receives the generic string "music" regardless of its valence/arousa…
All-time downloads
75
Public
Repo size
31.5 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors31.5 MB · 100%
At a glance
- Library
- peft
- License
- mit
- Access
- Public
- Created
- Jun 30, 2026
- Updated
- Aug 7, 2026
- SHA
- 01605917
Base models
Model card
From the Hugging Face model README
AffectScore -- LoRA r=32, no-affect (ablation)
Ablation checkpoint for the AffectScore rank sweep. Identical to the full r=32 model except Layer 1 V-A conditioning is removed: every training clip receives the generic string "music" regardless of its valence/arousal annotation. The LoRA adapter structure and Layer 2 style-text descriptor are otherwise unchanged.
This checkpoint represents the null baseline -- a LoRA adapter fine-tuned without any emotional conditioning. It is not intended for production use.
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 | 32 |
| LoRA targets | cross_attn.q_proj, cross_attn.v_proj |
| Layer 1 (V-A mood words) | disabled (replaced with "music") |
| Layer 2 (style-text descriptor) | enabled |
| Training clips | 1,571 (CC0/CC-BY, CLAP-quality filtered) |
| Epochs | 50 (best loss at epoch 30; mild overfitting thereafter) |
| Learning rate | 1e-4 |
| Batch size | 8 |
| Hardware | NVIDIA A100 (Google Colab) |
Training dataset: https://zenodo.org/records/21830658
Evaluation
All metrics computed on 200 held-out clips at 40 diffusion steps.
Audio quality (EVAL-02)
| FAD-MERT ↓ | CLAP-score ↑ | KLD ↓ | PCE ↑ |
|---|---|---|---|
| 29.13 | 0.32 | 1.23 | 3.57 |
Emotion controllability (EVAL-03)
| r_valence | RMSE_valence | r_arousal | RMSE_arousal |
|---|---|---|---|
| 0.005 | 0.486 | −0.012 | 0.308 |
Valence r=0.005, arousal r=−0.012 confirms the null baseline: without V-A conditioning, music2emo-predicted emotion is uncorrelated with designer intent on either axis.
Ablation variants
See affectscore-ace-step-r32-20260629 for the full variant table and evaluation discussion.
Citation
Citation will be added once the paper is published. If you use this model before then, please link to this repository.
- Library
- peft
- License
- mit
- Created
- Jun 30, 2026
- Updated
- Aug 7, 2026