langset-sounds-like

langset-sounds-like

A langset model: it reads an album review and emits a latent into a "how it sounds" geometry, so you can retrieve albums that sound alike — not albums with similar reviews. The review isn't a description of the sound; the model reads it and re-projects into the sound axis.

• Backbone: HuggingFaceTB/SmolLM2-135M + LoRA
• Latent: 576-d, in the backbone's own hidden space (your own embedding, not a re-projected off-the-shelf one)
• Objective: native self-contrastive — emit(review) is trained to land where emit(sonic-fingerprint) lands
• The geometry is the target text. These fingerprints weight vocals + emotion + instrumentation, so the space clusters by vocal register (yell-sing vs growl vs croon) and emotional tenor, not just instrumental heaviness. Rewrite the target text and the same recipe re-clusters by whatever you describe.

Trained with langset==0.2.0 on Hanno-Labs/sounds-like-fingerprints — album metadata + the sounds_like fingerprints used as target_text (no review text). input_text = the full album review; target_text = the album's sounds_like fingerprint.

Usage

pip install langset

from huggingface_hub import snapshot_download
from langset import LangSetModel

model = LangSetModel.load(snapshot_download("Hanno-Labs/langset-sounds-like"))
z = model.encode(["a wall of downtuned fuzz that buries the vocals under sheer volume"])
print(z.shape)   # (1, 576)

It's also a drop-in SetFit body:

body = model.as_sentence_transformer()          # a sentence_transformers.SentenceTransformer