oreva/bass_music_benchmark

BASS: Benchmarking Audio LMs for Musical Structure and Semantic Reasoning

BASS is a benchmark for evaluating music understanding and reasoning in audio language models. It comprises 2,658 questions across 12 tasks and 4 categories, covering 1,993 unique songs and over 138 hours of music.

---

🚀 Usage

from datasets import load_dataset

ds = load_dataset("oreva/bass_music_benchmark", "lyrics_transcription")
ds = load_dataset("oreva/bass_music_benchmark", "artist_collaboration")
ds = load_dataset("oreva/bass_music_benchmark", "structural_segmentation")
ds = load_dataset("oreva/bass_music_benchmark", "musicological_analysis")

---

🗂️ Task Categories

Category	Description	Tasks
Structural Lyrics Transcription	Transcribe song lyrics at varying levels of granularity	Full, Section-level
Structural Segmentation	Identify and localize structural sections within a recording	Full, Section-level
Artist Collaboration	Reason about the presence, duration, and identity of contributing artists	Counting, Duration, Localization, Attribution
Musicological Analysis	Assess knowledge of genre attributes across detection and ranking tasks	Single-Gene Detection, Pairwise Detection, Dominance Ranking, Attribution

---

📥 Audio Access

Due to copyright restrictions, audio files are not distributed directly. Each example includes a youtube_url field pointing to the source track on YouTube.

---

📬 Contact

For questions or feedback, please open a GitHub issue or reach out to us at minjang@cs.washington.edu and oahia@cs.washington.edu.

---

✍️ Citation

@article{jang2025bass,
  title   = {BASS: Benchmarking Audio LMs for Musical Structure and Semantic Reasoning},
  author  = {Min Jang and Orevaghene Ahia and Nazif Tamer and Sachin Kumar and Yulia Tsvetkov and Noah A. Smith},
  year    = {2025},
  url     = {https://arxiv.org/abs/2602.04085}
}