Where R2's data comes from

This page covers ensomi R2 1.0 and 1.0.1, not R1 or any earlier model. It summarises our data provenance report, which has the details, and claims nothing beyond it.

What R2 used, and what for

R2 was built from osu!'s ranked and loved 4K charts from 2 to 6 stars. We do not distribute these charts or their audio.

• Training. 11,368 eligible charts trained the rows model, which learns where notes start from the music and the mappers' timing, and R2, which arranges the notes without hearing the audio.
• Corpus prior. It takes a new chart's character from similar training charts. The published prior.json keeps one row of numbers, without names, for each training chart.
• Calibration. The published playability calibration and scope table are derived from the training charts.
• Timing. The mappers' tempo, beat grids and notes were references for choosing how R2 fits a beat grid to audio.
• Evaluation. Some charts were references for measuring and listening to generated charts.
• Previews. The four published previews use four source songs' audio. Their source mappers, named in the report, did not author the generated charts.
• Beat tracking. R2 uses BeatThis final0 as its authors trained it; the musicians in its training data are not fully recorded.

R2 is not built to imitate any particular mapper or artist: its models receive no mapper or artist names or IDs, were trained from scratch on the training charts as one pool, and no training draw, fine-tune, playability limit or evaluation measure is set per mapper or artist (evidence and limits).

The numbers, and what they are not

Counts of the eligible training charts on 2026-10-17. Labels are metadata claims, not verified people or rights holders.

11,368 charts

One chart is one difficulty, one .osu file.

1,129 Creator labels

Mapper names as written. Aliases are not merged or collaborations split, and guest mappers may be missing.

2,445 Artist labels

Artist names as written, not a list of composers, performers or rights holders.

5,026 audio hashes

Audio files by their exact bytes, not songs.

4,167 groups

Groups that keep related sources in the same data partition, not songs.

Eligible means in the pool that training drew from; whether or how often a chart was drawn is not recorded. The frozen inventory of 22,066 charts that the audit started from is not the training population.

What the audit shows, and what it does not

The audit establishes attribution and stage membership: which charts and labels each recorded stage of R2 used. It is not a rights clearance or a complete review of acquisition and permissions. It leaves open who is behind each label, which charts training actually drew, and the full list of musicians in BeatThis's training data.

What we decided

Policy statement of 2026-10-21

The released models r2-1.0 and r2-1.0.1 remain available with this disclosure. Future released models will use data covered by permissive licences or explicit permission suitable for the intended training and release uses; their data review considers, as applicable, inherited weights, upstream components, distillation and derived data. The audit does not show that R2's data meets this policy.

Enrollment and opt-out

We are exploring enrollment (opt-in) and opt-out, including a convenient way to verify that a decision comes from the rights holder. Nothing is in place yet, and we do not claim to remove what a released model has already learned.

Check the tables

stage_counts.csv gives each stage's counts, and creator_labels_by_stage.csv and artist_labels_by_stage.csv list every Creator and Artist label in each stage; their README explains them. To find a name, filter stage (r2_cache_fit_train for the training charts) and search raw_label, also for part of a name. Stages overlap: never add their counts.

Corrections and questions

To correct an attribution, ask how a chart or song was used, or say you are interested in enrollment or opt-out, open an issue on GitHub, in English, Japanese or Chinese.

Issues are public: include no personal data or audio. An issue does not by itself remove anything from a released model.