UniVerseSet

The training split of UniVerse (同谣).
Held-out evaluation lives in UniVerseBench.

UniVerseSet is the post-training corpus for large audio–language models on world folk music: ASR, captions, and audio-grounded chat, plus automatically transcribed ABC scores.

「诗言志，歌永言，声依永，律和声。」—《尚书·舜典》

Sister dataset (benchmark)	universe-team/universebench
Live museum demo	http://143.89.224.8:8790/

---

What's here

Archives (download and unpack; this repo is not a streaming load_dataset table — the viewer is disabled on purpose):

File	Role	Unpacked size / count
lyrics.tar.gz	lyrics transcribed by Qwen3-ASR; each file is {youtube_id}.txt	~173k files
captions.tar.gz	captions generated by Qwen3-Omni-Captioner; each file is {youtube_id}.txt	~152k files
dialogues.tar.gz	Multi-turn audio–chat generated by Qwen3-Next-80B-A3B-Instruct; each file is {youtube_id}.txt	~115k files
audio_ids.txt	Audio id; each line is youtube id	~174k lines
abc.tar.gz	Customly-improved SheetSage ABC transcriptions; each file is {youtube_id}.abc	~174k .abc files

The public chat line used in our experiments is the default SFT thinking set (~113k examples after filtering). Derived views (language-reweighted SFT, text DPO, multimodal audio-contrast DPO) are built from chat; see the UniVerse code repo training/data_prep/.

Intended use. Post-train LALMs so they listen to folk recordings rather than answering from language priors. The language mix is naturally skewed; our recipes keep that mix and apply language-weighted CE, text DPO (chosen vs rejected text, fixed audio), and audio DPO (positive vs negative audio, shared response).

---

Chat format (after unpacking dialogues.tar.gz)

{
  "messages": [
    {"role": "user", "content": "<audio>…"},
    {"role": "assistant", "content": "<think>\n…\n</think>\n\n…"}
  ],
  "audios": ["…"]
}

Index clips with audio_ids.txt. ABC files in abc.tar.gz are SheetSage transcriptions named by YouTube id (abc/<id>.abc; header + percussion / chord / melody voices).

---

How to download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="universe-team/universeset",
    repo_type="dataset",
    local_dir="universeset",
)
# then: tar -xzf dialogues.tar.gz  (etc.)

For evaluation, use UniVerseBench (material + qa).

---

License & attribution

Academic research only. Commercial use is prohibited.

All materials in UniVerseSet — including lyrics, captions, dialogues, ABC transcriptions, audio_ids.txt, this dataset card, and any derived annotations — are released solely for non-commercial academic research. You may not use the dataset, in whole or in part, to train, evaluate, or deploy models for commercial products, paid services, advertising, or other for-profit purposes.

Underlying folk recordings and YouTube-sourced audio remain under their original copyrights, collection licenses, and platform terms. This repository does not grant rights to redistribute or commercially exploit the audio. Please credit archives, collectors, and performers when citing or presenting clips.

License on the Hub: CC BY-NC 4.0. If you use UniVerseSet in research, please cite the paper below.

---

Citation

@article{universe2026,
  title   = {UniVerse: Benchmarking and Enhancing LALMs on Culturally Inclusive Low-resource Music Understanding},
  author  = {[TODO: authors]},
  journal = {arXiv preprint arXiv:[TODO]},
  year    = {2026}
}