Hugging Face's logo

Datasets:
ggmolly
/
ozut-data

Tasks:
Feature Extraction
Audio Classification
Modalities:
Audio
Tabular
Text
Formats:
parquet
Size:
100K - 1M
Tags:
osu
rhythm-games
beatmaps
audio
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
ozut-data / README.md
ggmolly's picture
ggmolly
Daily incremental dataset sync 2026-07-30
28c2982 verified
preview code
|
Raw
History Blame Contribute Delete
437 kB
metadata 
pretty_name: dataset used for ozut! research project
license: cc-by-nc-sa-4.0
task_categories:
  - feature-extraction
  - audio-classification
tags:
  - osu
  - rhythm-games
  - beatmaps
  - audio
configs:
  - config_name: beatmaps
    data_files:
      - split: full
        path: metadata/beatmaps.parquet
  - config_name: beatmapsets
    data_files:
      - split: full
        path: metadata/beatmapsets.parquet

dataset used for ozut! research project

Parsed osu! beatmaps, normalized metadata, hit objects, timing/control points, difficulty attributes, individual Opus audio, and raw .osu content in Parquet.

Table of contents

Snapshot

  • API beatmaps: 225,006
  • Parsed .osu files: 222,680
  • Parsed hit objects: 145,551,301
  • Artifact files: 282,288
  • Artifact payload: 180.78 GiB
  • Individual Opus files: 59,608
  • Approximate Hub files: 59,626
  • Parser: rosu-map 0.2.1
  • Difficulty calculator: rosu-pp 4.0.1

Exact source hashes and generation metadata are in snapshot.json. Integrity and coverage checks are in reports/validation.json.

Loading metadata 

from datasets import load_dataset

beatmaps = load_dataset("ozut-data", "beatmaps", split="full")
beatmapsets = load_dataset("ozut-data", "beatmapsets", split="full")

The normalized high-volume tables can be loaded directly as Parquet:

from datasets import load_dataset

hit_objects = load_dataset(
    "parquet",
    data_files={"full": "hf://datasets/ozut-data/metadata/hit_objects.parquet"},
    split="full",
)

map_key is the stable relation key between parsed beatmaps and these tables.

Artifact layout

Audio is stored as individual Opus files so a rightsholder request can be handled by deleting one file without rebuilding a container:

audio/<bucket>/<beatmapset_id>/<filename>.opus

The exact original .osu bytes are stored as binary values in metadata/osu_files.parquet. metadata/files.parquet maps each source artifact to either its individual repository path or its Parquet map_key.

Creating splits

This dataset deliberately provides the full corpus without predefined training, validation, or test sets. Create task-specific splits at beatmapset_id level so all difficulties from a beatmapset remain together. For experiments where song overlap would leak information, additionally group audio by the SHA-256 in metadata/files.parquet before assigning groups to splits.

Tables

  • metadata/beatmapsets.parquet: one row per API beatmapset.
  • metadata/beatmaps.parquet: one row per API beatmap, joined to the primary parsed file.
  • metadata/parsed_beatmaps.parquet: one row per parsed .osu file, including legacy files.
  • metadata/osu_files.parquet: exact original .osu bytes keyed by map_key.
  • metadata/difficulty_attributes.parquet: NM, HD, HR, DT, and HDDT stars/max combo.
  • metadata/break_periods.parquet: parsed breaks.
  • metadata/control_points.parquet: normalized timing, difficulty, effect, and sample points.
  • metadata/hit_objects.parquet: circles, sliders, spinners, and mania holds.
  • metadata/slider_control_points.parquet: slider path geometry.
  • metadata/files.parquet: artifact manifest and checksums.

The original optimized .osu bytes are retained in Parquet, so derived tables can be regenerated when parser or difficulty algorithms change.

Known coverage details

  • 21,783 legacy files with non-positive BeatmapID were matched unambiguously by beatmapset, mode, and difficulty version.
  • 168 parsed files could not be matched to the API snapshot and remain available in metadata/parsed_beatmaps.parquet and reports/unmatched_parsed_beatmaps.parquet.
  • 25 positive beatmap IDs occur in more than one source file. The deterministic primary candidate is used in beatmaps.parquet; every candidate remains in parsed_beatmaps.parquet.
  • 1 file header has a beatmapset ID different from its source artifact directory.
  • Parser failures: 0.
  • Missing resolved audio references: 0.

License

The original database compilation, schema, documentation, and maintainer-created transformations are licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license. See LICENSE for the full legal code.

This license applies only to rights the dataset maintainer is legally able to grant. It does not relicense third-party beatmaps, original .osu content, audio recordings, or source metadata. Those materials remain subject to their respective rightsholders and applicable terms.

Copyright and takedown policy

This research dataset includes the Opus audio referenced by the beatmaps. Metadata, beatmap files, and audio may have different rightsholders. Inclusion does not transfer ownership, imply endorsement, or grant rights beyond those held by the original authors and rightsholders.

Rights holders may request removal through the Hugging Face repository's Community tab. A request should identify the affected track, beatmap, or beatmapset and the requester's relationship to the work. On the first good-faith request from a rightsholder, the affected audio file and its manifest row will be removed in the next dataset revision. No fee is required.

Credits

This dataset features work from:

Counts refer to parsed beatmap difficulty files. Profile links come from the corresponding osu! beatmap metadata.