STEMMA-Instruct

📄 Paper  |  💻 Code  |  🤗 STEMMA-Bench  |  🤗 STEMMA-MF2601  |  🤗 STEMMA-MOSS-Music

Multi-audio music instruction data from STEMMA: Song-to-Stem Multi-Audio Reasoning for Large Audio Language Models, built from JamendoMaxCaps and the commercially reusable subset of FMA. The in-house dataset used in the paper is not included.

Instances (main)	162,832 = Relation MultiQA 81,416 + Coreference 81,416
Songs	JamendoMaxCaps 69,297 · FMA 12,119
Audio inputs per instance	2–5 (songs, sections, or section stems)

Format

Field	Description
id	instance ID
task	<relation>_<format> (29 relation tasks) or coreference
question	prompt with one <audio> per entry of audios
answer	target
audios	audio IDs

Configs: main, relation_only, coreference_only. Relation MultiQA formats are yes/no, short answer, and four-candidate multiple choice; Coreference describes 2–4 songs under separate labels.

Audio

Audio is not included. Audio IDs map to the original tracks (Jamendo / FMA track IDs): <corpus>/masters/<track>/<track>-NNN-<section> (section), <corpus>/blocks/<track>/<section>/...-<role> (SCNet stem), <corpus>/masters/<track> (full song). Section boundaries are in metadata/<corpus>.duckdb, which also holds track, clip, and tag metadata; captions/ holds the full-song captions used by Coreference.

Citation

@misc{sohn2026stemma,
  title         = {{STEMMA}: Song-to-Stem Multi-Audio Reasoning for Large Audio Language Models},
  author        = {Sohn, Hoyeol and Kim, Wonil and Kim, Keunhyoung and Kum, Sangeun and Kim, Taehyoung and Moon, Dongjoo and Charoenchob, Theerasak and Weerapang, Teeratep and Lee, Jongpil and Nam, Juhan},
  year          = {2026},
  eprint        = {2610.11884},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2610.11884}
}

License

CC BY-NC-SA 4.0. The original audio and metadata remain subject to the JamendoMaxCaps and FMA terms.