sleeping-ai/SUNO-1M

SUNO 1 Million

Update 04/07/2025

Excited to share, SUNO Dataset, the largest known collection of artificially generated songs from the most popular AI music generator SUNO. SUNO dataset is an on-going effort of Sleeping AI to provide groundbreaking and impactful datasets to create and research for state-of-art Generative Music modelling tasks. It is a collection of datasets that includes Sleeping-DISCO, SUNO and many more.

Why SUNO?

SUNO is one of the key three players in the market, backed by a wonderful research team and community that loves to make songs. Artificially generated songs are still heavily debated and majority of share in the research is held by Chinese universities and labs. While their quality is debatable and I myself never believed this can be a business so fast and soon.

Hey! We are here now and I am sharing this dataset with everyone.

How the dataset was sourced and collected?

I won't disclose the actual mechanism behind its collection but we did not scrape the website because cloudflare and SUNO's paid Clerk tokens are notoriously obnoxious to get past.

To quote:

It was sourced responsibly and maybe if SUNO team did a better job in their engineering, I won't be able to download it.

Data fields

uuid: a unique identification for each audio

user_id: to trace each user and where the data coming-from

audio_url: LINK to audio

video_url: LINK to video

model_name: Model used for creating the song

major_model_version: Tracking model version across songs

How about a DMCA takedown?

Yes, it is possible but there are many other versions of SUNO's database on Huggingface and Kaggle that are-up. I am obviously not the first. Besides, I allow the option if you think there was a very very original song that's present in this dataset, I will take it down. Just point me to its uuid

Btw, I followed all the standard data collection rules that's mentioned in EU, I only scraped the data that was publicly available and I did it purely for scientific research and purpose. Besides, SUNO probably trained its models on someone else's data as well.

+I will highlight that there's a paper called

1. SUNOCAPs

This papers have used SUNO songs before. I don't believe my dataset did anything different than what many others have did before.

Fine, Why its better?

Stark difference between Me and my competition will be, I have deduplicated and validated each data-point I have uploaded. Previously, nyuuzyou did a nice contribution of 655K songs but it was riddled with jsonl garbage that exposed personal information and unnecessary context. And it did not provide estimate of entire SUNO database and failed provide a valid user list. Mine does following:

1. We are the first high-quality SUNO dataset to reach 1M (million) songs.
2. We offer both audio and video in nicely done format.
3. We have plans to provide additional data by burning GPU hours for the sake of research and a nice arXiv paper.
4. We estimate SUNO has 1.5-2M songs. and we had a 33% error rate considering 1.5M being database size.

Ethics statement

We have compiled the data in good faith and made sure that we respected individual privacy and we have downloaded the data under our local and European law.

LICENCE

We offer this dataset under CC-by-nc-nd 4.0, which prohibits following acts:

1. Data can't be shared and individuals aside from the original author is not allowed to create derivatives.
2. You have to seek permission of Sleeping AI, before modifying and sharing online.

Feel free to reach out on X (formerly twitter): https://x.com/sleeping4cat