Bring Your Own Music: try the generated benchmark yourself
What you will see
20 of the 300 multiple-choice items that MusICA-MetaBench generated from the ChoraleBricks dataset. For each one you get the same piece and the same question that eight multimodal LLMs were given.
Why these 20
We picked them by hand to be diverse:
- items that all models answered correctly, items that no model did, and items where the models disagree;
- the same question about the same piece as an audio recording, as ABC notation and as a score image, so you can compare the modalities;
- different kinds of questions: pitch, interval, rhythm, temporal proportion and harmony.
How it works
- Listen, read or look, and pick one of the five options.
- See the ground truth and what each model answered, with its raw response.
- Open “Why this item was selected” to see what the item shows.
This page stores and sends nothing; your answers are gone when you reload it. See a problem? Write us an email (sourada@ufal.mff.cuni.cz). Paper: arXiv:2607.06015, code: GitHub.
Answer each item, then compare with what eight multimodal LLMs answered.
Why this item was selected
See a problem with this item or with how a model's answer was evaluated? Write us an email (sourada@ufal.mff.cuni.cz).
Thank you for trying the demo
We would like to hear from you, about anything from a single item to working together:
- Question formulation: was a question unclear, ambiguous or misleading? Write us.
- Question selection: are these 20 items a good showcase, are the skills we test the right ones, and which perceptual skills are we missing? Write us.
- Collaboration: would you add a dataset of your own music, or a question template with its ground-truth extractor? Write us, or see how to contribute.
If the links do not open your mail client, write to sourada@ufal.mff.cuni.cz.
Bring your own music.