Add dataset card

README.md

@@ -1,3 +1,74 @@
 ---
 license: apache-2.0
+viewer: false
 ---
+
+# 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.
+
+The training split has 10,000 items (2,500 per task). The test split has 1,000 clean items (250 per task) and
+2,000 FLIP/STAY items (1,000 FLIP and 1,000 STAY).
+
+Code, scoring and the generator: https://github.com/AMAAI-Lab/MusicListenBench/tree/release
+Leaderboard: https://amaai-lab.github.io/MusicListenBench/
+
+## Files
+
+| File | Content |
+|---|---|
+| `mlb_test_clean.tar.gz` | 1,000 clean test files |
+| `mlb_test_flip_stay.tar.gz` | 2,000 FLIP/STAY test files |
+| `mlb_train_melody.tar.gz`, `mlb_train_harmony.tar.gz`, `mlb_train_timbre.tar.gz`, `mlb_train_rhythm.tar.gz` | 2,500 training files each |
+| `SHA256SUMS` | SHA-256 sums of the six archives |
+| `data/eval.json` | the 1,000 clean test items |
+| `data/eval_perturb.json` | the 2,000 FLIP/STAY test items |
+| `data/eval_perturb.meta.jsonl` | for each FLIP/STAY item: condition, what was changed, base answer |
+| `data/train.json` | the 10,000 training items |
+
+The archives are 4.0 GB in total. The test archives alone are 0.9 GB.
+
+## Audio
+
+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. The clean and training files unpack to
+`audio/<task>/` and the FLIP/STAY files to `audio_perturb/<task>/`, with the tasks `Q1_melody`, `Q2_harmony`,
+`Q3_timbre` and `Q4_rhythm`.
+
+```bash
+hf download amaai-lab/MusicListenBench --repo-type dataset --local-dir musiclistenbench
+cd musiclistenbench
+sha256sum -c SHA256SUMS
+for f in *.tar.gz; do tar -xzf "$f"; done
+```
+
+With the code repository, `bash scripts/fetch_audio.sh https://huggingface.co/datasets/amaai-lab/MusicListenBench/resolve/main`
+does the same and puts the audio under `data/`; add `--test-only` to skip the training archives.
+
+## Items
+
+The item files are lists of records:
+
+```json
+{"voice": ["Q1_melody/Q1_melody__test__100__000000__base.wav"],
+ "conversations": [{"from": "human", "value": "Listen carefully to the two melodic clips, ... <audio>"},
+                   {"from": "gpt", "value": "B"}]}
+```
+
+`voice` is the audio path below `audio/` (`audio_perturb/` for `eval_perturb.json`), the first conversation turn is
+the question and the second is the correct answer, `A` or `B`. Every question ends with an instruction that says
+which letter means which answer; the meaning differs between question wordings. The file name without `.wav` is
+the item ID. FLIP files end in `__flip__<edit>` (`note_identity`, `third_quality`, `instrument`, `pulse_rate`)
+and STAY files in `__stay__<effect>` (`eq`, `hiss`, `room`, `transposition`).
+
+## 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.