Sparse Steerable Retrieval — BatchTopK SAEs
BatchTopK SAEs trained on MuQ-MuLan music4all audio embeddings (512-d), one per sparsity level L0. Each L0-*/ holds last.ckpt (encoder/decoder weights) + config.yaml. Load via steerable_retrieval.steer.load_steerable_sae(repo, subfolder='L0-20').
