YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Audio-Audio CLAP SAE (32x expansion, k=5)

Sparse Autoencoder trained on 155M CLAP embeddings from 6 audio datasets.

Architecture

  • Input dimension: 512
  • Hidden dimension: 16384 (32x expansion)
  • Top-k: 5
  • Alive features: 14,127 / 16,384
  • Dead features: 2,257

Training Data

  • MLS, CommonVoice, AudioSnippets, Maestrino, Emolia, Podcast
  • Total samples: 86,729,608

Usage

from sae import SparseAutoencoder
sae = SparseAutoencoder.load_from_disk("path/to/model")
latents = sae.encode(embeddings)  # (batch, 512) -> (batch, 16384)

Stats

  • Total feature firings: 433,648,040
  • Mean firings per sample: 5.0
Downloads last month
29
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support