STEMMA-Bench

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

Multi-audio music relation benchmark from STEMMA: Song-to-Stem Multi-Audio Reasoning for Large Audio Language Models. Each item has an anchor (Audio A) and four candidates (Candidate B–E), one correct (chance 25%). Clips are sections and native stems of multitrack recordings; every item was checked by listening.

Items	300 (100 each from MUSDB18-HQ, MoisesDB, MedleyDB)
Structural / attribute	127 / 173
Tasks	14

Format

Field	Description
id	item ID
task	<relation>_anchor_candidates_mcq
question	prompt with five <audio> (Audio A, Candidate B–E)
answer	correct letter (B–E)
audios	audio IDs, in prompt order

Structural tasks: same_track, same_track_block, adjacent_master, adjacent_block, same_section_cross_role, parent_child; the rest are attribute tasks (key, tempo, instrument, genre). _block tasks use stems. Accuracy is reported per group; predicting A counts as wrong.

Audio

Audio is not included. Audio IDs are <corpus>/<track>/<section>[/<stem>]; metadata/audio.jsonl gives each clip's source files in the original corpus, start/end time, and whether summed stems are peak-normalized. 210 MedleyDB section mixes come from the mixture files in seungheondoh/cmd-medleydb-metadata (revision 99e4ae2).

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 audio remains subject to the MUSDB18-HQ, MoisesDB, and MedleyDB terms.