Shorten dataset card

README.md

@@ -12,8 +12,6 @@ tags:
 - music
 - multi-audio
 - music-qa
-- source-separation
-- audio-language-model
 configs:
 - config_name: main
   data_files: data/main.jsonl
@@ -36,7 +34,7 @@ dataset_info:
     sequence: string
 ---
 
-# STEMMA-Instruct: Multi-Audio Instruction Data over Songs, Sections, and Stems
+# STEMMA-Instruct
 
 <p align="center">
   <a href="https://github.com/hoyso48/STEMMA">💻 Code</a> &nbsp;|&nbsp;
@@ -45,98 +43,35 @@ dataset_info:
   <a href="https://huggingface.co/hoyso48/STEMMA-MOSS-Music">🤗 STEMMA-MOSS-Music</a>
 </p>
 
-**STEMMA-Instruct** is the instruction-tuning set of *STEMMA: Song-to-Stem Multi-Audio Reasoning for Large Audio
-Language Models*. Each instance gives a model two to five music clips — full songs, sections, or section stems — and
-asks about their relations. Labels come from catalog provenance (which song, section, and stem a clip belongs to),
-not from a language model.
+Multi-audio music instruction data from *STEMMA: Song-to-Stem Multi-Audio Reasoning for Large Audio Language Models*,
+built from JamendoMaxCaps and the commercially reusable subset of FMA. The in-house dataset used in the paper is not
+included.
 
-This release covers the public corpora JamendoMaxCaps and the commercially reusable subset of the Free Music Archive
-(FMA). The in-house dataset used in the paper is not included. The release is text and metadata only; audio is rebuilt
-from the original corpora.
-
-## Key Statistics
-
-| Statistic | Value |
+| | |
 |---|---|
-| **Instances** (`main`) | 162,832 |
-| **Relation MultiQA** | 81,416 |
-| **Coreference** | 81,416 |
-| **Songs** | 81,416 (JamendoMaxCaps 69,297 · FMA 12,119) |
-| **Audio inputs per instance** | 2–5 |
-| **Relation tasks** | 29 |
-
-## Dataset Structure
-
-| Field | Type | Description |
-|---|---|---|
-| `id` | string | Instance identifier (e.g., `fma-relation-000007`) |
-| `task` | string | Relation task (`<relation>_<format>`) or `coreference` |
-| `question` | string | Prompt with one `<audio>` placeholder per entry of `audios`, in order |
-| `answer` | string | Target response |
-| `audios` | list[string] | Audio IDs (see [Audio](#audio)) |
-
-| Config | Rows | Content |
-|---|---:|---|
-| `main` | 162,832 | Relation MultiQA + Coreference |
-| `relation_only` | 81,416 | Relation MultiQA |
-| `coreference_only` | 81,416 | Coreference |
+| Instances (`main`) | 162,832 = Relation MultiQA 81,416 + Coreference 81,416 |
+| Songs | JamendoMaxCaps 69,297 · FMA 12,119 |
+| Audio inputs per instance | 2–5 (songs, sections, or section stems) |
 
-## Task Types
+## Format
 
-**Relation MultiQA** (one instance per song) asks about a relation between clips. Three formats:
-
-- **pair_yes_no** (34,466): two clips, yes/no.
-  *"Are Audio A and Audio B different isolated components from the same recorded section?"* → `yes`
-- **pair_short_answer** (18,812): two clips, a short answer listing shared keys or tags.
-  *"What genres do Audio A and Audio B share?"* → `The genres Audio A and Audio B share are folk, pop, singersongwriter.`
-- **anchor_candidates_mcq** (28,138): an anchor and four candidates, one correct.
-  *"Which candidate audio comes from the same track as Audio A?"* → `C`
-
-**Coreference** (81,416) gives two to four full songs under separate labels and asks for a description of each; the
-target repeats each song's own caption under its label.
-
-| Group | Relations |
+| Field | Description |
 |---|---|
-| Structural | same track, adjacent section, same section (different stem), parent–child (section → stem), same stem type |
-| Attribute | same key, same tempo, faster tempo, instrument overlap, genre overlap |
+| `id` | instance ID |
+| `task` | `<relation>_<format>` (29 relation tasks) or `coreference` |
+| `question` | prompt with one `<audio>` per entry of `audios` |
+| `answer` | target |
+| `audios` | audio IDs |
 
-Relations are asked over sections, over stems, or across the two (parent–child). Task names without `_block` use
-sections; `_block` tasks use stems.
+Configs: `main`, `relation_only`, `coreference_only`. Relation MultiQA formats are yes/no, short answer, and
+four-candidate multiple choice; Coreference describes 2–4 songs under separate labels.
 
 ## Audio
 
-Audio IDs use the original track IDs (Jamendo track ID; FMA track ID zero-padded to six digits):
-
-| Audio ID | Clip |
-|---|---|
-| `<corpus>/masters/<track>/<track>-NNN-<section>` | one section of the song |
-| `<corpus>/blocks/<track>/<section>/<track>-NNN-<section>-<role>` | one stem of that section (`vocals`, `drums`, `bass`, `other`) |
-| `<corpus>/masters/<track>` | the full song (its sections in order); used by Coreference |
-
-Section boundaries are in `metadata/<corpus>.duckdb`; stems were separated with SCNet. The
-[code repository](https://github.com/hoyso48/STEMMA) rebuilds the clips from the original audio.
-
-## Metadata and Captions
-
-| File | Content |
-|---|---|
-| `metadata/{jamendomaxcaps,fma}.duckdb` | `tracks` (title, duration, key, tempo), `sections` (boundaries), `samples` (clip unit, stem role, key, tempo, quality flag), `genres`, `instruments`, `chords`, `moods` |
-| `captions/{jamendomaxcaps,fma}.jsonl` | `{"audio": "<corpus>/masters/<track>", "caption": "..."}`, the full-song captions used by Coreference |
-
-## Usage
-
-```python
-from datasets import load_dataset
-import duckdb
-
-data = load_dataset("hoyso48/STEMMA-Instruct", "main", split="train")
-example = data[0]
-print(example["task"], example["question"], example["answer"], example["audios"], sep="\n")
-
-# Section boundaries for a clip (after downloading metadata/fma.duckdb)
-con = duckdb.connect("metadata/fma.duckdb", read_only=True)
-print(con.execute("select * from sections where track_id = '086595' order by section_index").fetchall())
-```
+Audio is not included. Audio IDs map to the original tracks (Jamendo / FMA track IDs):
+`<corpus>/masters/<track>/<track>-NNN-<section>` (section), `<corpus>/blocks/<track>/<section>/...-<role>` (SCNet stem),
+`<corpus>/masters/<track>` (full song). Section boundaries are in `metadata/<corpus>.duckdb`, which also holds track,
+clip, and tag metadata; `captions/` holds the full-song captions used by Coreference.
 
 ## Citation
 
@@ -151,5 +86,4 @@ print(con.execute("select * from sections where track_id = '086595' order by sec
 
 ## License
 
-**CC BY-NC-SA 4.0.** The original audio and metadata remain subject to the terms of JamendoMaxCaps and the Free Music
-Archive.
+CC BY-NC-SA 4.0. The original audio and metadata remain subject to the JamendoMaxCaps and FMA terms.