---
license: apache-2.0
configs:
- config_name: clean
  default: true
  data_files:
  - split: train
    path: data/train/**
  - split: test
    path: data/test/**
- config_name: flip_stay
  data_files:
  - split: test
    path: data/test_flip_stay/**
---

# MusicListenBench

Audio and item files for *MusicListenBench: Can Audio LLMs Hear the Difference Yet?*

Each test item is one audio file (clip A, one second of silence, clip B) and one question about the pair: same
melody, same chord, same instrument, or which clip is faster. Every item has two variants. In the FLIP variant
the music changes, so the correct answer changes. In the STAY variant only the sound changes (room echo, EQ,
background hiss, transposition), so the correct answer stays. A model that ignores the audio scores 50% on
average, whatever letter it prefers.

Code, scoring and the generator: https://github.com/AMAAI-Lab/MusicListenBench/tree/release
Leaderboard: https://amaai-lab.github.io/MusicListenBench/

## Splits

| Config | Split | Files | Folder |
|---|---|---|---|
| `clean` | `train` | 10,000 (2,500 per task) | `data/train/` |
| `clean` | `test` | 1,000 (250 per task) | `data/test/` |
| `flip_stay` | `test` | 2,000 (500 per task: 250 FLIP and 250 STAY) | `data/test_flip_stay/` |

Each folder holds the audio in one sub-folder per task (`Q1_melody`, `Q2_harmony`, `Q3_timbre`, `Q4_rhythm`) and a
`metadata.jsonl`. Audio is 16-bit PCM WAV, 44.1 kHz, mono. A file is clip A, one second of silence and clip B, so
it is 6.0 s long for melody and timbre, 5.0 s for harmony and 9.0 s for rhythm.

```python
from datasets import load_dataset
train = load_dataset("amaai-lab/MusicListenBench", "clean", split="train")
test = load_dataset("amaai-lab/MusicListenBench", "clean", split="test")
flip_stay = load_dataset("amaai-lab/MusicListenBench", "flip_stay", split="test")
```

## Columns

* `audio`: the file
* `item_id`: the file name without `.wav`
* `task`: `Q1_melody`, `Q2_harmony`, `Q3_timbre` or `Q4_rhythm`
* `question`: the question that is asked with the audio
* `answer`: the correct answer, `A` or `B`
* `condition`, `factor` (`flip_stay` only): `flip` or `stay`, and what was changed. FLIP edits are `note_identity`,
  `third_quality`, `instrument` and `pulse_rate`; STAY effects are `eq`, `hiss`, `room` and `transposition`.

Every question ends with an instruction that says which letter means which answer; the meaning differs between
question wordings.

## Item files

`data/train.json`, `data/eval.json` (the clean test items), `data/eval_perturb.json` (the FLIP/STAY test items) and
`data/eval_perturb.meta.jsonl` are the item files of the code repository. They list the same items as the
metadata, with the question ending in the `<audio>` placeholder. The audio paths in them are relative to
`data/train/`, `data/test/` and `data/test_flip_stay/`.

## Licences

The audio is rendered with the soundfonts FluidR3 GM, GeneralUser GS and MuseScore General; follow the licence
terms that come with them. The benchmark audio contains no recordings of people and no copyrighted musical
works.