leeyujun/Beyond-Binary-Instrument-QA

🎵 Beyond Binary Instrument QA:
Probing Instrument Grounding in Music Audio-Language Models

Yujun Lee · Joonhyeok Shin · Hyoeun Kim · Kyuhong Shim
Sungkyunkwan University
📄 arXiv   |   🤗 Dataset

Benchmark release. Five complementary evaluation configurations test instrument presence, reduced genre-prior reliance, fine-grained discrimination, long-context multi-label recognition, and temporal localization. The release contains 15,580 test examples with OpenMIC-2018-derived audio.

---

Contents

• Benchmark Introduction
• Configurations
• OpenMIC-2018
• Preview Subsets
• Citation
• License

---

Benchmark Introduction

Beyond Binary Instrument QA is a benchmark suite for evaluating whether music audio-language models ground their answers in the instruments that are actually audible. It extends beyond standard binary presence questions with controlled prior reduction, candidate-based instrument discrimination, 30-second multi-label reasoning, and segment-level temporal localization.

This archive-oriented release provides the canonical benchmark metadata and evaluation audio for five test-only configurations. Each configuration contains one complete metadata.jsonl, one TAR.GZ archive with its full audio/ directory, and a small preview subset for inspection. Model predictions, experiment code, notebooks, API outputs, and credentials are not included.

Property	Value
Total examples	15,580
Configurations	5
Audio duration	10 seconds or 30 seconds
Split	test
Source audio	OpenMIC-2018
Distribution	metadata JSONL + one audio TAR.GZ per configuration

Configurations

Configuration	Evaluation focus	Duration	Examples	Audio archive
binary_presence	Binary instrument-presence QA	10 s	9,332	binary_presence_audio.tar.gz
prior_reduced	Genre-prior-reduced binary QA	10 s	590	prior_reduced_audio.tar.gz
instrument_discrimination	Candidate-based instrument discrimination	10 s	1,051	instrument_discrimination_audio.tar.gz
long_context_multilabel	Long-context multi-label recognition	30 s	1,028	long_context_multilabel_audio.tar.gz
temporal_localization	Three-segment temporal localization	30 s	3,579	temporal_localization_audio.tar.gz

Each full metadata record contains:

• audio_archive: the archive filename in the same configuration directory.
• audio_member: the exact path of that example inside the archive, beginning with audio/.

Task-specific fields—including the original question, candidate order, source sample keys, and gold answer or answers—are preserved from the benchmark metadata used in the accompanying experiments.

OpenMIC-2018

All released audio is derived from OpenMIC-2018, an open dataset for multiple-instrument recognition. The source clips and original instrument annotations were made available by Spotify AB under CC BY 4.0.

• OpenMIC-2018 on Hugging Face
• Authoritative OpenMIC-2018 release
• OpenMIC-2018 paper

Gold labels use the benchmark's high-confidence OpenMIC relevance annotations. The 10-second configurations use OpenMIC-derived clips. The 30-second long-context and temporal-localization configurations concatenate three OpenMIC-derived 10-second segments in the source order recorded by source_sample_keys.

Preview Subsets

Each configuration includes preview/metadata.jsonl and exactly 20 corresponding individual audio files under preview/audio/. The examples were selected reproducibly with seed 42 and remain present in the complete archive.

The previews are intended for quickly inspecting the schema and listening to representative examples. They are not full evaluation splits and should not be used to reproduce benchmark scores.

Preview selection balances binary answers, instrument groups, long-context group and answer-set diversity, or temporal target ranges as appropriate for each configuration.

Citation

If you use this benchmark, please cite both the Beyond Binary Instrument QA paper and OpenMIC-2018.

@article{lee2026beyond,
  title={Beyond Binary Instrument QA: Probing Instrument Grounding in Music Audio-Language Models},
  author={Lee, Yujun and Shin, Joonhyeok and Kim, Hyoeun and Shim, Kyuhong},
  journal={arXiv preprint arXiv:2606.31338},
  year={2026}
}

@inproceedings{humphrey2018openmic,
  title={Openmic-2018: An open data-set for multiple instrument recognition.},
  author={Humphrey, Eric and Durand, Simon and McFee, Brian},
  booktitle={ISMIR},
  pages={438--444},
  year={2018}
}

License

This release is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

The source audio clips and original instrument annotations are derived from OpenMIC-2018, which is made available by Spotify AB under CC BY 4.0. Beyond Binary Instrument QA selects, reorganizes, and annotates OpenMIC-2018-derived material. The long-context and temporal-localization configurations additionally concatenate three OpenMIC-derived 10-second segments into 30-second evaluation examples.

When sharing or adapting this material, users must provide appropriate attribution, retain the license notice, and indicate whether changes were made. See the repository's LICENSE file and the full CC BY 4.0 license.