Datasets:

---

ProjectAquarius

/

colour-regions

Like 1

Follow

Project Aquarius 4

Adds 37 recordings in two new rounds: round 4 (22, a Northeast Plain playlist) and round 5 (15, a Beijing playlist). Derived data only; no audio. The same rows are in branch metadata-rounds-4-5 of github.com/huiwenn/aquarius (metadata/), so the two stay in step.

What changes

• data/recordings.parquet 1,655 -> 1,692 rows; data/song_labels.parquet 1,469 -> 1,497 songs (28 new, 8 existing songs gain recordings); data/fingerprints.parquet and data/link_status.parquet +37 rows each (links checked 2026-10-23, all alive).
• 220 new files: midi/{primary,game,rosvot}/, musicxml/primary/, f0/{rmvpe,pesto}/ (midi/game is missing for 2 of the new items, as for 12 existing ones).
• Card numbers updated (recordings, hours, YouTube count, songs, subsets, series count); rounds 4 and 5 added to the round list.
• The four tables have exactly the same schema as before. The 1,655 existing recording rows are unchanged, and so are the existing song rows, apart from the item_ids and n_recordings of the 8 songs that gained recordings.

Dedupe

• 38 candidates were checked by video ID and by Chromaprint fingerprint (bit error rate <= 0.25 = same recording) against all 1,655 existing items.
• One duplicate removed: YRA4TggquC4 is a re-upload of By334DK4svU (CCTV, round 1), BER 0.036.
• The other 37 are not in the dataset: no ID overlap; nearest fingerprint BER 0.39 or higher.

How the derived columns were made

• group_id: channel_singer_groups from build_release.py on the whole corpus; all 1,655 existing ids come out unchanged.
• core / core_exclusion / subset: the rules of add_core (composed, folk-style composed, speech). No new exclusion reasons.
• Transcriptions: the same pipeline as rounds 2-3 (separation, GAME/ROSVOT/YourMT3+, ensemble, RMVPE and PESTO F0).
• The tables were extended directly because the team data/ folder was not available. A rebuild from that folder must add rounds 4 and 5, or it will drop these rows.

Please review (hand-made or consequential)

1. 28 new song labels are written from my own knowledge or from uploader descriptions: song_label_check = unchecked, confidence = low, ids NOR-1469 to NOR-1496. Five credits (顾冠仁; 郑路 and 马洪业 with arranger 阎惠昌; 小楼听雨) come from uploader descriptions only. The song-id registry (song_ids.csv) on the team drive is not updated and must add these ids.
2. 17 of the 37 are song_class = opera (9 full 二人转 pieces, 5 Peking opera clips, 3 percussion ensembles). By the current rules they are core, like the 3 existing opera rows. Please decide whether that is intended. Of the rest, 3 are composed (excluded) and 1 is folk_style_composed.
3. 16 of the 22 round-4 items are from the Musical Map of China channel. Round 3 was built from a catalogue of eligible items; please check these were not left out on purpose.
4. anthology_overlap and score_matches were not rechecked for the new songs.

Noticed in the existing dataset (not changed here)

• main is older than GitHub's metadata/ for some existing rows: 1 recording is in Gan here but in Hakka on GitHub, and 164 provenance tiers, 35 song_label_check values and 24 subset values differ. A rebuild with build_hf.py refreshes them.
• The card sentence "regions are unequal in size (73-202 recordings)" may be stale: all rows give 60-202.

Updated after review: the recording lists of the 8 changed song rows now use the same ; separator as the other rows, and an earlier note about song rows with mismatched item lists was a mistake in my check and is removed.

Generated with Claude Code (https://claude.com/claude-code)