thejorseman/CloneHeroDatasetCharts

Clone Hero Charts Dataset

Dataset Description

Tokenized Clone Hero charts with beat-level audio conditioning. Each row is one instrument track (guitar / bass / drums) from one song.

Feature	Value
Total rows (train)	43,665
Parquet shards	1753
Audio: MERT embeddings	Yes [num_beats, 768]
Audio: log-mel frames	Yes [num_beats, 32, 128]

Dataset Structure

Data Fields

Column	Type	Description
song_id	string	MD5 hash of artist + title
instrument	string	"guitar" / "bass" / "drums"
source_format	string	"chart" or "midi"
tokens	list[int32]	Token sequence (model target)
num_tokens	int32	Token count
num_beats	int32	Beat count
mert_embeddings	list[list[float32]]	MERT embeddings per beat [B, 768]
logmel_frames	list[list[list[float32]]]	Log-mel spectrogram [B, 32, 128]
beat_times_s	list[float32]	Beat start times (seconds)
beat_durations_s	list[float32]	Beat durations (seconds)
bpm_at_beat	list[float32]	BPM at each beat
time_sig_num_at_beat	list[int32]	Time signature numerator
time_sig_den_at_beat	list[int32]	Time signature denominator
song_name	string	Song title
artist	string	Artist name
genre	string	Genre tag
charter	string	Charter name
year	int32	Release year
song_length_ms	int32	Song length in milliseconds
difficulty	int32	Difficulty 0–6 (−1 = unset)
resolution	int32	Tick resolution (normalised to 192)
has_star_power	bool	Track contains star-power sections
has_solo	bool	Track contains solo sections
has_dedicated_stem	bool	Instrument-specific audio stem available
num_notes	int32	Total note count
notes_per_beat_mean	float32	Average notes per beat
chord_ratio	float32	Fraction of notes that are chords
sustain_mean_ticks	float32	Mean sustain length in ticks
bpm_mean	float32	Mean BPM
bpm_std	float32	BPM standard deviation

Data Splits

Split	Rows
train	43,665

Usage

from datasets import load_dataset

# Stream individual shards (low RAM):
ds = load_dataset("thejorseman/CloneHeroDatasetCharts", split="train", streaming=True)
for row in ds:
    print(row["song_name"], row["instrument"], len(row["tokens"]))

Source Data

Charts scraped from public Clone Hero repositories in .chart and .mid format. Audio conditioning extracted with:

• MERT (m-a-p/MERT-v1-330M) — 768-dimensional embeddings, mean-pooled per beat
• Log-mel — 128-bin spectrograms resampled to 32 time frames per beat

License

MIT — see LICENSE.

Citation

@misc{clonecharter2025,
  title  = {Clone Hero Charts Dataset},
  author = {The Jorseman},
  year   = {2025},
  url    = {https://huggingface.co/datasets/thejorseman/CloneHeroDatasetCharts}
}