Shorten dataset card

README.md

@@ -11,7 +11,6 @@ task_categories:
 tags:
 - music
 - multi-audio
-- music-qa
 - benchmark
 configs:
 - config_name: default
@@ -35,7 +34,7 @@ dataset_info:
     num_examples: 300
 ---
 
-# STEMMA-Bench: A Multi-Audio Benchmark for Song, Section, and Stem Relations
+# STEMMA-Bench
 
 <p align="center">
   <a href="https://github.com/hoyso48/STEMMA">💻 Code</a> &nbsp;|&nbsp;
@@ -44,87 +43,37 @@ dataset_info:
   <a href="https://huggingface.co/hoyso48/STEMMA-MOSS-Music">🤗 STEMMA-MOSS-Music</a>
 </p>
 
-**STEMMA-Bench** is the evaluation set of *STEMMA: Song-to-Stem Multi-Audio Reasoning for Large Audio Language
-Models*. Each item gives an anchor clip (`Audio A`) and four candidates (`Candidate B`–`E`) and asks which candidate
-stands in a given relation to the anchor. Clips are sections and section stems of multitrack recordings, so stems are
-the recordings' own stems rather than separated ones. Every item was checked by listening. The benchmark shares no
-corpus with STEMMA-Instruct.
+Multi-audio music relation benchmark from *STEMMA: Song-to-Stem Multi-Audio Reasoning for Large Audio Language
+Models*. Each item has an anchor (`Audio A`) and four candidates (`Candidate B`–`E`), one correct (chance 25%). Clips
+are sections and native stems of multitrack recordings; every item was checked by listening.
 
-## Key Statistics
-
-| Statistic | Value |
+| | |
 |---|---|
-| **Items** | 300 (100 each from MUSDB18-HQ, MoisesDB, MedleyDB) |
-| **Structural / attribute items** | 127 / 173 |
-| **Tasks** | 14 |
-| **Songs** | 481 |
-| **Clips** | 1,466 |
-| **Audio inputs per item** | 5 (anchor + 4 candidates) |
-| **Chance level** | 25% (answers balanced over `B`–`E`) |
-
-## Dataset Structure
-
-| Field | Type | Description |
-|---|---|---|
-| `id` | string | Item identifier (`stemma-bench-000` … `stemma-bench-299`) |
-| `task` | string | Relation task (`<relation>_anchor_candidates_mcq`) |
-| `question` | string | Prompt with five `<audio>` placeholders: `Audio A`, then `Candidate B`–`E` |
-| `answer` | string | Correct candidate letter (`B`–`E`) |
-| `audios` | list[string] | Audio IDs in prompt order (see [Audio](#audio)) |
-
-Example:
+| Items | 300 (100 each from MUSDB18-HQ, MoisesDB, MedleyDB) |
+| Structural / attribute | 127 / 173 |
+| Tasks | 14 |
 
-> **Audio A**, **Candidate B–E** · *"Which candidate audio comes from the same track as Audio A? Answer only with the candidate letter."* → `E`
+## Format
 
-## Tasks
-
-| Group | Task | Items |
-|---|---|---:|
-| Structural | `same_track` (sections) · `same_track_block` (stems) | 22 · 21 |
-| | `adjacent_master` (sections) · `adjacent_block` (stems) | 21 · 21 |
-| | `same_section_cross_role` (another stem of the same section) | 21 |
-| | `parent_child` (the stem contained in the anchor section) | 21 |
-| Attribute | `same_key` · `same_key_block` | 22 · 21 |
-| | `same_tempo` · `same_tempo_block` | 22 · 21 |
-| | `faster_tempo` · `faster_tempo_block` | 22 · 21 |
-| | `instrument_overlap` · `genre_overlap` (sections) | 22 · 22 |
-
-Scoring: accuracy, reported separately for the structural and attribute groups. Predicting `A` (the anchor) counts as
-wrong.
-
-## Audio
-
-This dataset contains text and audio-reconstruction metadata only. Audio IDs are `<corpus>/<track>/<section>` for a
-section and `<corpus>/<track>/<section>/<stem>` for a stem, using each corpus's own track and stem identifiers.
-`metadata/audio.jsonl` lists, for every clip, its source files inside the original corpus, the start and end time, and
-whether the summed mixture is peak-normalized:
-
-```json
-{"audio_id": "medleydb/AimeeNorwich_Flying/sf_005/S04", "source": "medleydb",
- "files": ["V1/AimeeNorwich_Flying/AimeeNorwich_Flying_STEMS/AimeeNorwich_Flying_STEM_04.wav"],
- "start_sec": 116.6447, "end_sec": 139.0856, "peak_normalize": false}
-```
-
-| `source` | Corpus |
+| Field | Description |
 |---|---|
-| `musdb18hq` | MUSDB18-HQ |
-| `moisesdb` | MoisesDB v0.1 |
-| `medleydb` | MedleyDB V1/V2 |
-| `medleydb_mixture_mirror` | MedleyDB mixtures from [seungheondoh/cmd-medleydb-metadata](https://huggingface.co/datasets/seungheondoh/cmd-medleydb-metadata) (revision `99e4ae2`, `medleydb.tar.gz`) |
+| `id` | item ID |
+| `task` | `<relation>_anchor_candidates_mcq` |
+| `question` | prompt with five `<audio>` (Audio A, Candidate B–E) |
+| `answer` | correct letter (`B`–`E`) |
+| `audios` | audio IDs, in prompt order |
 
-The [code repository](https://github.com/hoyso48/STEMMA) rebuilds every clip sample for sample. 210 MedleyDB section
-mixes come from the mirror, which is a different, louder mix than the official `_MIX.wav`; use it to reproduce the
-reported results.
+Structural tasks: `same_track`, `same_track_block`, `adjacent_master`, `adjacent_block`, `same_section_cross_role`,
+`parent_child`; the rest are attribute tasks (key, tempo, instrument, genre). `_block` tasks use stems. Accuracy is
+reported per group; predicting `A` counts as wrong.
 
-## Usage
-
-```python
-from datasets import load_dataset
+## Audio
 
-bench = load_dataset("hoyso48/STEMMA-Bench", split="test")
-item = bench[0]
-print(item["task"], item["question"], item["answer"], item["audios"], sep="\n")
-```
+Audio is not included. Audio IDs are `<corpus>/<track>/<section>[/<stem>]`; `metadata/audio.jsonl` gives each clip's
+source files in the original corpus, start/end time, and whether summed stems are peak-normalized. 210 MedleyDB section
+mixes come from the mixture files in
+[seungheondoh/cmd-medleydb-metadata](https://huggingface.co/datasets/seungheondoh/cmd-medleydb-metadata)
+(revision `99e4ae2`), not the official `_MIX.wav`.
 
 ## Citation
 
@@ -139,4 +88,4 @@ print(item["task"], item["question"], item["answer"], item["audios"], sep="\n")
 
 ## License
 
-**CC BY-NC-SA 4.0.** The audio remains subject to the terms of MUSDB18-HQ, MoisesDB, and MedleyDB.
+CC BY-NC-SA 4.0. The audio remains subject to the MUSDB18-HQ, MoisesDB, and MedleyDB terms.