JacobLinCool/taiko-1000-parsed-clean

Taiko-1000 Parsed — Clean Split

A leakage-free, group-aware re-split of JacobLinCool/taiko-1000-parsed: ~1000 songs paired with their Taiko no Tatsujin charts (TJA + parsed per-course note structures), intended for training and evaluating music-to-chart generation models.

The content of the rows is identical to the original dataset (same audio, same tja, same parsed oni/hard/normal/easy/ura course structs, same metadata). This release differs only in how the rows are partitioned into splits and adds three provenance columns so that duplicate songs can never straddle a split boundary.

Why this dataset exists — the leakage that motivated it

A provenance audit of the original dataset found song-level train/test leakage. The original two-split layout (train = 1039 rows, test = 116 rows) was built at the chart-entry level, so the same song could appear in both splits. Three independent identity keys were checked per song and agreed exactly:

• metadata.TITLE (NFKC-normalized),
• audio.path (the per-song .ogg filename), and
• SHA-256 of the raw .ogg audio bytes (the strongest key).

25 of the 116 test songs (21.6%) also appear in the original train split. Title, audio path, and audio content-hash all give the same intersection of 25 with no ambiguity. For those songs, any model (or evaluation-time language model) trained on train had already seen the exact same audio+charts it was later asked to score on test. The audit also revealed intra-split duplicate audio: only 927 distinct audio hashes over the 1039 original train rows, and 113 over the 116 test rows — i.e. some songs contribute multiple chart entries with byte-identical audio.

This clean split removes both problems.

Deduplication and split design

• Grouping (union-find). All 1155 rows are grouped by song identity using a union-find over two edge types: (a) identical raw-audio SHA-256, and (b) identical NFKC-normalized TITLE. Two rows that share either key land in the same group. This yields 1013 song groups over the 1155 rows.
• Group-level splitting. Splits are assigned per group, never per row, so every version/chart-entry of a song stays together in one split. The split is balanced on ura (hidden-course) presence and BPM, with a fixed seed = 20260710.
• Multiple charts per song are kept. When a group contains several rows (e.g. multiple chart entries for the same audio), all rows are retained and the representative row of the group is flagged with canonical = True. Use canonical to deduplicate to one row per song when desired.

Verified guarantees

• Cross-split raw-audio SHA-256 overlap = 0 (train/validation/test, pairwise).
• Cross-split TITLE overlap = 0.
• Verified both at manifest-build time and by reloading the pushed dataset from the Hub.

Splits

Split	Song groups	Rows	ura fraction	BPM median
train	813	924	0.193	170.0
validation	80	91	0.200	177.6
test	120	140	0.192	173.0
total	1013	1155	—	—

(Genre metadata is unknown/absent for almost all songs in the source data.)

Columns

Same schema as the original dataset, plus three provenance columns:

• audio — Audio feature (44.1 kHz .ogg, decoded to a mono waveform).
• tja — raw TJA chart file text.
• oni, hard, normal, easy, ura — parsed per-course chart structs (course, level, player, balloons, and per-measure segments with timed notes). A course may be None if the song has no chart for it (ura is absent for most songs).
• metadata — TJA header fields (TITLE, BPM, OFFSET, GENRE, ...).
• group_id (int) — duplicate-group id from the union-find; rows sharing audio or title share a group_id.
• canonical (bool) — True for the one representative row of each group.
• audio_sha256 (string) — SHA-256 hex digest of the raw .ogg bytes; the identity key used for grouping and the leakage guarantee.

Usage

from datasets import load_dataset

ds = load_dataset("JacobLinCool/taiko-1000-parsed-clean")
# one row per unique song:
train_unique = ds["train"].filter(lambda r: r["canonical"])

Provenance / reproducibility

The exact per-row split assignment is derived from a clean-split manifest (clean_split_manifest.json, included in this repo) of the form {sid: {split, group_id, canonical, audio_sha256}}, where sid is "{original_split}_{row_index:05d}". During the build, the SHA-256 recomputed from each row's raw .ogg bytes was asserted to equal the manifest's audio_sha256 for every row before pushing.

License and intended use

This dataset inherits the terms of the source dataset JacobLinCool/taiko-1000-parsed: it is provided for research, non-commercial use only. Audio and chart copyrights belong to their respective owners. Do not redistribute the audio for commercial purposes.

Citation

If you use this dataset, please reference the source dataset and note that this is the leakage-free, group-aware split.