---
library_name: audio-interv
tags:
  - ace-step
  - audio
  - feature-selection
  - interpretability
  - music
  - sae
  - sparse-autoencoder
  - steering
  - vocal-gender
---

# SAE Feature-Selection Scores — `vocal_gender` (ACE-Step)

Per-concept feature-importance scores for the ACE-Step SAEs at `transformer_blocks.6.cross_attn` and `transformer_blocks.7.cross_attn`. Consumed at inference time by `SAESteeringController` via `load_features_from_score_cache` (top-k features per diffusion step).

## Files

- `tf7_scores.pkl` — scores for the tf7 SAE.
- `tf6_scores.pkl` — scores for the tf6 SAE.

Each pickle is a dict keyed by selection method (`tfidf`, `diff`, `mean_pos`, ...); values are tensors of shape `(num_timesteps, num_features)`.

## Paper

TADA! Tuning Audio Diffusion Models through Activation Steering — [https://huggingface.co/papers/2602.11910](https://huggingface.co/papers/2602.11910)

## Quickstart

```python
from src.steering.methods.sae import load_features_from_score_cache

top20_tf7 = load_features_from_score_cache(
    "lukasz-staniszewski/ace-step-sae-scores-vocal-gender", score_filename="tf7_scores.pkl", top_k=20,
)
top20_tf6 = load_features_from_score_cache(
    "lukasz-staniszewski/ace-step-sae-scores-vocal-gender", score_filename="tf6_scores.pkl", top_k=20,
)
```