vibebench / README.md
Divi19's picture
v1.2 release (16)
cba4f8f verified
|
Raw History Blame Contribute Delete
1.94 kB
metadata
license: cc-by-4.0
task_categories:
  - text-to-audio
  - audio-classification
tags:
  - music
  - retrieval
  - benchmark
configs:
  - config_name: queries
    data_files: data/queries.parquet
  - config_name: corpus
    data_files: data/corpus.parquet

VibeBench

A benchmark for searching music by vibe: 500 real music requests and 1,831 Creative Commons songs, with LLM-graded relevance.

  • data/queries.parquet: requests (first messages from CPCD), with the dev/test split used for Divi19/vibe-embed.
  • data/corpus.parquet: the songs, with artist, title, licence, source link and the written description the grader reads.
  • audio/: a 30-second excerpt of each song (MP3), as listed in corpus.parquet.
  • data/judgements.parquet, data/qrels.tsv: grades (0–3) from Phi-4; data/qrels_rescaled.tsv: the same grades rescaled with cut-points fitted on IncompeBench's human-validated grades (the main board).
  • scoreboard.md: results.

Grading caveat. The grader agrees with human-validated grades at κ ≈ 0.47, below the usual 0.6 bar. Use VibeBench as a secondary benchmark.

Licences

  • Requests: CPCD, the Conversational Playlist Curation Dataset (Google Research), CC BY 4.0.
  • Audio: each excerpt keeps its own licence (CC0 1.0, CC BY or public domain), listed per song with its artist and source link in corpus.parquet and CREDITS.csv. The excerpts are 30-second cuts of the original tracks.
  • Descriptions, grades and everything else here: CC BY 4.0.

VibeBench requests: first messages from CPCD, the Conversational Playlist Curation Dataset (Google Research), CC BY 4.0, https://github.com/google-research-datasets/cpcd (commit db633b012aa9607598cd635db9e7f7623e812767). Messages are used as typed (whitespace normalised); a subset was selected automatically. Judge check: IncompeBench (Mixedbread AI), https://huggingface.co/datasets/mixedbread-ai/incompebench-lenient