junhao1122/Classical-Chinese-Poetry-Songs

Classical Chinese Poetry Songs

Version 1.0 · Chinese poetry-to-song dataset · MIT License

This is the frozen dataset used in Classical Chinese Poetry Song Generation: A Curated Dataset and Domain Adaptation (working manuscript title). It contains synthetic songs with vocals and accompaniment, poetic lyrics, audio-grounded captions, original generation descriptions, and work-level splits. The upstream song model was identified by generation providers as Suno V6; this label was not independently verified.

Contents and fixed partitions

Partition	Works	Samples	Hours
Train	613	1,170	32.3743
Validation	78	150	4.1234
Test	75	143	3.7842
Total	766	1,463	40.2820

The 1,463 samples represent 1,457 source recordings: 1,451 complete-song MP3s and 12 WAV segments from six longer recordings. The six unsegmented parent files are not duplicated in this training-format release; their identities, hashes, and segment frame ranges are retained. No audio is transcoded or normalized during packaging. The twelve segments were already part of the frozen experiment, not newly cut for release.

All versions and segments of a work stay in the same partition. These are the experiment's frozen partitions, not a fresh random split. The freeze retained 75 hash-ordered works from the earlier test pool and moved the one overflow work to validation; it did not select works using scores. This does not claim that all historical test material was never accessed.

Muse-style packaging

The distribution layout follows Muse: flat, uncompressed .tar audio archives and separate JSONL metadata with relative paths. Our smaller corpus uses one archive per partition rather than many language-specific shards.

README.md
train.jsonl
validation.jsonl
test.jsonl
train_audio.tar
validation_audio.tar
test_audio.tar
release_manifest.json
SHA256SUMS
LICENSE

Download the versioned dataset with the Hugging Face CLI:

hf download junhao1122/Classical-Chinese-Poetry-Songs \
  --repo-type dataset --revision v1.0 --local-dir Classical-Chinese-Poetry-Songs

Verify and extract from the downloaded dataset directory:

sha256sum -c SHA256SUMS
for split in train validation test; do
  mkdir -p "audio/$split"
  tar -xf "${split}_audio.tar" -C "audio/$split"
done

Each archive contains only unique sample_id.mp3 or sample_id.wav filenames. Metadata points to paths such as audio/train/catalog_002a740c1456__0.mp3. Do not rename extracted audio independently of its metadata.

Read records without additional dependencies:

import json
from pathlib import Path

root = Path(".")  # Dataset directory, after extraction.
rows = [json.loads(line) for line in (root / "train.jsonl").read_text().splitlines()]
audio_path = root / rows[0]["audio_path"]
assert audio_path.is_file()
print(rows[0]["lyrics"], rows[0]["caption"])

Hugging Face can display the JSONL metadata. The audio_path column is a string, not embedded audio: downloading/loading the metadata does not extract the tar archives or automatically decode songs.

Annotation fields

• sample_id, poem_id, split, track_index: stable sample and work identities.
• audio_path, audio_sha256, duration: relative audio filename, SHA-256 of distributed bytes, and frozen duration in seconds.
• title, author, text: literary title, historical author, and full source work text. text is not a segment transcript; use lyrics for supervision.
• lyrics, lyrics_sha256: exact sample lyrics, including explicit repetition and structural tags, and SHA-256 of the UTF-8 string. Historical textual variants and any inherited markup are retained, not silently corrected.
• caption, caption_source: adopted audio-grounded training description and annotation provenance. view_media denotes model-assisted listening, not a human label; parent_view_media denotes inheritance by a segment.
• original_caption, generation_parameters: original synthesis description and requested settings. Requested BPM/key/meter are not measured audio labels.
• generation_group: with_reference used descriptions informed by existing performances; no_reference and without_reference name two historical no-reference batches. The original categories are not silently merged. There are 277 performance-informed samples; no human performance audio is included.
• provider: archived service-provider value, or null when absent. Missing history is not inferred from today's default provider.
• source: allowlisted literary source URL, title, anchor, revision and hash where recorded. Internal cache paths and reference-audio file paths are omitted.
• lyric_admission: recorded curation route, not an exhaustive pronunciation certification. Segment admission is inherited from the parent plus alignment.
• parent_sample_id, parent_audio_sha256, parent_lyrics_sha256: provenance for segments; null for full-song samples.
• source_start_frame, source_end_frame, source_frames, sample_rate, alignment_method: locally derived segment boundaries. Start is inclusive, end exclusive, in decoded parent frames. These are not manually corrected word timestamps or provider-returned timing annotations.
• language, is_instrumental: fixed sample attributes.

Optional fields are present as null where unavailable. Raw API responses, temporary download URLs, request identifiers, account details, local paths, human reference audio, model weights, and evaluation outputs are not included.

Construction and limitations

Poetic lyrics preserve the selected historical text. Works of at most 64 Chinese characters repeat as verse and chorus; longer works are used once. Original music descriptions and later audio-grounded captions are kept separate. Curation combines provider-returned lyric comparison, ASR diagnostics, model-assisted listening, selected author checks, audio integrity checks, and caption similarity. Long songs are split at locally aligned lyric boundaries.

This is not a human-recorded singing corpus or an exhaustive manual pronunciation audit. Similarity and automatic quality scores do not establish caption factual accuracy, human preference, copyright clearance, or freedom from pretraining overlap. Segment captions can describe the whole parent song. The raw parent ASR CER and provider lyric text are deliberately omitted rather than presented as segment-level ground truth. Fine-grained style sections and word timestamps from Muse are not claimed for this dataset.

License

This dataset is released under the MIT License, as authorized by the publisher. Retain the license notice and source metadata when redistributing. The dataset is not affiliated with or endorsed by Suno or its service providers.

Citation

@misc{classicalchinesepoetrysongs2026,
  title = {Classical Chinese Poetry Songs},
  author = {{Classical Chinese Poetry Songs contributors}},
  year = {2026},
  version = {1.0},
  howpublished = {Hugging Face dataset},
  url = {https://huggingface.co/datasets/junhao1122/Classical-Chinese-Poetry-Songs}
}

For reproducible use, download the v1.0 revision and verify SHA256SUMS. Questions and correction requests can be submitted through this dataset repository's Community tab. Include the relevant sample_id when reporting a data issue.