universe-team/universe_qa_check

UniVerse · 同谣

One shared verse — across cultures, minds, and ages.

UniVerse (同谣) is a multilingual folk-music resource for large audio–language models (LALMs). It has two complementary parts:

Split	Role
UniVerseSet	Post-training corpus: audio-grounded conversations over world folk recordings
UniVerseBench	Adversarial multiple-choice benchmark for folk-music understanding

Together they cover Chinese folk (regional styles and minority languages), Korean gugak, and world folk in 38+ languages, with curated metadata, scores, and bilingual notes.

「诗言志，歌永言，声依永，律和声。」—《尚书·舜典》

---

What's in this hub

This Hugging Face dataset hosts the audio + annotations for both splits. Training and evaluation code lives in the UniVerse GitHub repository. An interactive sound museum is available as a live demo.

	UniVerseSet	UniVerseBench
Purpose	SFT / DPO / latent post-training	Held-out evaluation
Scale	~113k audio–chat examples (v3 “11w-think” line)	5,042 MCQ items
Audio	Training folk recordings (YouTube-sourced + curated collections)	Benchmark clips: Chinese / Korean / world
Format	Chat messages + audios (ms-swift / Qwen-Omni style)	{id, audios, question, options, answer}
Typical use	Language-balanced SFT, text DPO, audio-contrast (M2) DPO, REPA latent phases	Strict-parser accuracy by language group and question type

Headline stats used on the project site: 1.4M metadata tokens, 38+ languages, 5,042 QA, 1,080 minutes of benchmark audio.

---

UniVerseSet (training)

UniVerseSet is the instruction / preference corpus used to post-train Qwen2.5-Omni and Qwen3-Omni on folk music.

Each example pairs one or more audio clips with a multi-turn dialogue. Assistant turns may include a <think>…</think> block before the final answer, matching the “thinking” SFT format used in our recipes.

Typical fields:

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

The public line used in our experiments is the v3.0 thinking set (~113,015 examples). Derived views (language-reweighted SFT, text DPO, multimodal audio-contrast DPO) are built from this set; see the GitHub training/data_prep/ scripts.

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

---

UniVerseBench (benchmark)

UniVerseBench is a multilingual, adversarial multiple-choice benchmark. Items were selected so that strong LALMs cannot rely on surface cues: we keep questions that models miss, Gemini optimization logs flag as misleading, and we enforce coverage per recording.

• 5,042 items in the final release (universe_benchmark_v1_all_final.jsonl)
• Museum / metadata backbone: 372 exhibits — 102 Chinese, 90 Korean, 180 world folk
• Most items use one clip; a subset uses two or three clips for comparison

Question types

Code	Name
FE	Feature Extraction
CR	Causal Reasoning
PR	Music Pattern Recognition
CA	Comparison Analysis
ED	Error Detection

Language groups (for reporting): Chinese (zh), Korean (ko), other languages (L-oth.).

Audio quality mix (benchmark recordings): indoor professional, indoor non-professional, and low-quality field recordings.

Example:

{
  "id": "adversarial_000001",
  "audios": ["audios/benchmark/chinese/0000.mp3"],
  "question": "<audio>…",
  "options": {"A": "…", "B": "…", "C": "…", "D": "…"},
  "answer": "A"
}

Evaluate with the scripts under eval/ in the code repo (strict option-letter parser; thinking models use the answer after </think>).

---

Collections at a glance

Chinese folk. Provinces and ethnic minorities, with ABC / MusicXML scores and native + English notes where available.

Korean gugak. Arirang lineages and vernacular songs, with province and tori (melodic mode) metadata.

World folk. Dozens of languages under one protocol — from Arabic muwashshah to Japanese sakura songs — spanning families such as Indo-European, Uralic, Afro-Asiatic, Turkic, Japonic, and others.

---

How to load

from datasets import load_dataset

train = load_dataset("<org>/UniVerse", "UniVerseSet", split="train")
bench = load_dataset("<org>/UniVerse", "UniVerseBench", split="test")

Replace <org>/UniVerse with the actual hub id after upload. Audio paths inside the JSONL are relative to the dataset root (audios/...).

---

License & attribution

Code and cards: Apache-2.0. Folk recordings remain under their original source licenses / collection terms; see per-file metadata. Please credit regional archives and performers when redistributing clips.

---

Citation

@article{universe2026,
  title   = {UniVerse: Multilingual Adversarial Folk-Music Understanding for Large Audio-Language Models},
  author  = {[TODO: authors]},
  journal = {arXiv preprint arXiv:[TODO]},
  year    = {2026}
}