ACE-Step SAE feature scores — mood

Per-timestep feature-importance score tables for the ACE-Step tf6/tf7 cross-attention SAEs: activations collected with the shared CONCEPT_TO_PROMPTS pairs (the same text every data-driven method in the benchmark uses) and encoded with the SAE in fp32. Provenance (seed, git SHA) in the *_scores_config.json sidecars.

ctrl = SAESteeringController.from_pretrained(
    "lukasz-staniszewski/ace-step-sae-scores-mood", hookpoint="transformer_blocks.7.cross_attn",
    scores_cache_path="lukasz-staniszewski/ace-step-sae-scores-mood", score_filename="tf7_scores.pkl",
    top_k=20, intervention_mode="steering_vector", alpha=10.0,
)

The original paper-era score tables remain available at revision 231f2763eded80426e3d68ba7db81f0ac63b8aef of this repo (revision="231f2763eded80426e3d68ba7db81f0ac63b8aef" in hf_hub_download).

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including lukasz-staniszewski/ace-step-sae-scores-mood

TADA Steering Collection

Collection

Steering vectors, SAEs, and counterfactual prompts for ACE-Step. Companion to the audio-interv unified Controller interface. • 87 items • Updated Sep 10 • 2