Archit00/bbench-dep-cmi-bench

CMI-Bench: A Comprehensive Benchmark for Evaluating Music Instruction Following

Authors: Yinghao Ma, Siyou Li, Juntao Yu, Emmanouil Benetos, Akira Maezawa

🎉🎉🎉 Paper accepted by the 26th conference of the International Society for Music Information Retrieval (ISMIR). See you in Daejeon, Korea from September 21-25, 2025.

---

Abstract

Recent advances in audio-text large language models (LLMs) have opened new possibilities for music understanding and generation. However, existing benchmarks are limited in scope, often relying on simplified tasks or multi-choice evaluations that fail to reflect the complexity of real-world music analysis. We introduce CMI-Bench, a comprehensive music instruction-following benchmark designed to evaluate audio-text LLMs on a diverse set of music information retrieval (MIR) tasks. CMI-Bench reinterprets a broad range of traditional MIR annotations into an instruction-following format and uses standardized evaluation metrics consistent with state-of-the-art MIR models. Our experiments reveal significant performance gaps between current LLMs and specialized supervised models, as well as cultural, chronological, and gender biases. CMI-Bench establishes a unified foundation for evaluating and advancing music-aware LLMs.

🚀 Key Contributions

• Comprehensive Task Coverage: CMI-Bench includes 14 diverse MIR tasks, moving beyond simple classification to include regression, captioning, and complex sequential tasks.
• Standardized Evaluation: Unlike previous benchmarks that rely on multiple-choice questions, CMI-Bench employs open-ended, task-specific metrics aligned with the MIR literature (e.g., using mir_eval), allowing for direct comparison with traditional supervised models.
• Evaluation Toolkit: We provide a full evaluation toolkit that supports all major open-source audio-textual LLMs, enabling standardized and reproducible benchmarking.
• In-depth Analysis: The benchmark facilitates a deeper analysis of model capabilities, including generalization, prompt sensitivity, and biases related to culture and gender.

🎵 Tasks and Datasets

CMI-Bench encompasses 14 tasks evaluated across 20 different datasets, covering a wide range of challenges in music information retrieval.

Task	Dataset(s)	Metric(s)
Genre Classification	MTG-Genre, GTZAN	ROC-AUC, PR-AUC, Accuracy
Emotion Tagging	MTG-Emotion	ROC-AUC, PR-AUC
Emotion Regression	EMO	$R^2$
Instrument Classification	MTG-Instrument, Nsynth-Instrument	ROC-AUC, PR-AUC, Accuracy
Music Tagging	MagnaTagATune, MTG-Top50	ROC-AUC, PR-AUC
Pitch Estimation	Nsynth-Pitch	Accuracy
Key Detection	GiantSteps	Gmean Score
Lyrics Transcription	DSing	WER, CER
Music Captioning	SDD, MusicCaps	BLEU, METEOR, ROUGE, Bert-Score
Melody Extraction	MedleyDB v2	Melody Accuracy
(Down)Beat Tracking	GTZAN-Rhythm, Ballroom	F-measure
Vocal Technique	VocalSet	Accuracy
Performance Technique	GuZheng 99	Frame-level micro/macro-F1

This is a summary of the tasks listed in Table 1 of the paper.

🤖 Models Evaluated

Here is a revised version of the README section that improves clarity, structure, and consistency with the accompanying table:

---

Evaluated Models

We benchmark 11 publicly available audio-text large language models (LLMs), representing a diverse range of architectures and training paradigms. These models vary in scale, input modality coverage (sound, speech, music), and design choices across encoders and decoders.

A summary of each evaluated model’s capabilities is shown below:

Model	#Params	Sound	Music	Speech
Pengi	323M	✓	✓	✗
Audio-Flamingo	2.2B	✓	✓	✗
LTU	7B	✓	✓	✗
LTU-AS	7B	✓	✓	✓
MusiLingo-long	7B	✗	✓	✗
MuLLaMA	7B	✗	✓	✗
GAMA	7B	✓	✓	✗
GAMA-IT	7B	✓	✓	✗
Qwen-Audio-Chat	8.4B	✓	✗	✗
Qwen2-Audio-Instruct	8.4B	✓	✓	✓
SALMONN-Audio	13B	✓	✓	✓

Note: "Sound" refers to general non-speech audio; "Music" and "Speech" indicate support for those modalities in both input understanding and reasoning tasks.

📊 Key Findings

1. LLMs Underperform Supervised Baselines: Across most tasks, instruction-following LLMs fall significantly short of task-specific supervised MIR models, except in music captioning.
2. Generalization is Limited: Models perform best on datasets that were likely part of their training corpus, indicating that generalization to unseen or structurally different tasks remains a key challenge.
3. Sequential Tasks are Challenging: All models struggle with tasks requiring structured, time-based outputs like melody extraction and beat tracking. This is likely due to the diversity in prompt formats and limited exposure to dense temporal supervision during training.
4. Emotion Regression Fails: No model provides usable predictions for arousal and valence, highlighting a fundamental gap in mapping continuous perceptual attributes from music.
5. Cultural and Gender Bias: A fine-grained analysis reveals biases toward Western instruments and pop genres. We also observed performance differences in identifying male versus female voices.

🛠️ Getting Started with the Toolkit

The CMI-Bench evaluation toolkit is designed for easy and standardized evaluation of audio-text LLMs on MIR tasks. This section guides you through preparing datasets, running inference with audio-text LLMs, and evaluating results using the CMI-Bench toolkit.

🛠️ 0. Installation

To install model-specific environments (e.g., Qwen-audio, Qwen2-audio, Audio-Flamingo, Mu-LLaMA, MusiLingo, LTU, LTU-AS), please refer to:

📄 CMI-bench/model/README.md

Each model has its own setup instructions and pre-trained checkpoints.

🛠️ 1. Prepare the Dataset

🛠️ 1.1 Download Test Audio

Download test-set audio from Hugging Face:

wget https://huggingface.co/datasets/nicolaus625/CMI-bench/resolve/main/test_Data.zip
unzip test_Data.zip -d CMI-bench/data

🛠️ 1.2 Generate JSONL Annotation Files

To create instruction-following data pairs in .jsonl format:

# Example: Generate beat tracking data
python CMI-bench/data/Beat-Transformer/sft_beat.py

This creates files like:

CMI-bench/data/Beat-Transformer/CMI_ballroom_beat.jsonl

Repeat similarly for other tasks by running sft_*.py scripts in CMI-bench/data/*/.

🛠️ 2. Inference the Model

Run inference using:

python model/infer.py \
  --model qwen2 \
  --output-file results

This command will:

• Load the specified model
• Process each input audio and instruction under ~/CMI-bench/data/*/CMI*.jsonl
• Save predictions to model/results/{model}/{model}_{task}.jsonl

Available models: qwen, qwen2, salmonn, musilingo, ltu, ltu_as, mullama, flamingo, etc.

🛠️ 3. Configure Your Own Model

To add your own model:

1. Extend infer.py with a new --model option.

2. Implement a get_{model_name}_pred() function that takes:

   • text (instruction)
   • audio_path (test audio path)
   • any required processors or tokenizers

3. Place output JSONL results in model/results/{model}/.

🛠️ 4. Run Evaluation

To evaluate model outputs using task-specific metrics:

python evaluate.py \
  --model qwen2 \
  --task ballroom_beat

You can replace --task with:

• A specific dataset (e.g., GTZAN, MusicCaps, MTG_emotion)
• Or --task all to run evaluation for all available tasks

Results include metrics like:

• ROC-AUC / PR-AUC (for multi-label tasks)
• WER / CER (for lyrics transcription)
• Accuracy (for multi-class classification )
• R² (for emotion regression)
• F1 (for structured outputs like beat tracking or techinique detection)
• BLEU / BERTScore (for music captioning)

📜 Citation

If you use CMI-Bench in your research, please cite our paper:

@misc{ma2025cmibenchcomprehensivebenchmarkevaluating,
      title={CMI-Bench: A Comprehensive Benchmark for Evaluating Music Instruction Following}, 
      author={Yinghao Ma and Siyou Li and Juntao Yu and Emmanouil Benetos and Akira Maezawa},
      year={2025},
      eprint={2506.12285},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2506.12285}, 
}

License

This work is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0).