FreeTZAN

🤗 Dataset: huggingface.co/datasets/gclef-cmu/freetzan (download, or browse and listen in the dataset viewer)

FreeTZAN is a small, openly licensed music genre classification dataset in the spirit of GTZAN: 1,200 30-second clips in 10 balanced genres, with fixed train / valid / test splits (1000 / 100 / 100) that share no artists. All audio comes from the Free Music Archive (FMA) under Creative Commons, CC0, or Public Domain Mark licenses.

• It is a toy dataset for simple, pedagogical MIR demonstrations, not a rigorous benchmark. Unlike GTZAN, it can be redistributed, and every clip comes with attribution and license metadata.
• The labels are noisy. They are the genre tags that FMA users gave their own tracks, taken at face value without manual checking. Some clips are mislabeled, and 100% test accuracy is likely impossible.

Classes

Each class is a set of FMA genres (a genre id and its subtree in FMA's genres.csv):

class	train	test	FMA genres
rock	100	10	Rock (12) without Punk (25), plus Power-Pop (111)
punk	100	10	Punk (25) without Power-Pop (111)
electronic	100	10	Electronic (15) without Ambient Electronic (42)
ambient	100	10	Ambient (107), Ambient Electronic (42)
hiphop	100	10	Hip-Hop (21)
folk	100	10	Folk (17)
pop	100	10	Pop (10)
experimental	100	10	Experimental (38)
classical	100	10	Classical (5)
jazz	100	10	Jazz (4)

Quick start

Download the dataset, then train a minimal MFCC + SVM classifier (takes a few minutes):

pip install huggingface_hub librosa scikit-learn pandas
hf download gclef-cmu/freetzan --repo-type dataset --local-dir freetzan
cd freetzan

import librosa, numpy as np, pandas as pd
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.svm import SVC

def mfcc_stats(path):
    y, sr = librosa.load(path)
    m = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=20)
    return np.concatenate([m.mean(axis=1), m.std(axis=1)])

md = pd.read_csv('metadata.csv')  # one row per clip
X = np.stack([mfcc_stats(p) for p in md.path])
train, test = md.split == 'train', md.split == 'test'
clf = make_pipeline(StandardScaler(), SVC()).fit(X[train], md.genre[train])
print('test accuracy:', clf.score(X[test], md.genre[test]))  # 0.39

The Hub copy also loads with datasets.load_dataset('gclef-cmu/freetzan'), with splits train, validation, and test.

Baselines

scikit-learn classifiers on clip-level features. Features are standardized, models are trained on train, and each classifier's one hyperparameter is picked on valid. Test accuracy (chance is 10%):

Features	Dim	kNN	LogReg	SVM (RBF)	Random forest
MFCC mean/std	80	36%	38%	44%	43%
FMA librosa features	518	42%	40%	52%	45%
LAION-CLAP (music)	512	61%	66%	65%	62%
MERT-v1-95M	768	49%	60%	59%	52%

• MFCC mean/std: 20 MFCCs and their deltas (librosa), mean and standard deviation over time.
• FMA librosa features: FMA's precomputed features.csv.
• LAION-CLAP (music): music_audioset_epoch_15_esc_90.14.pt via the laion_clap package, averaged over three 10 s windows.
• MERT-v1-95M: hidden states averaged over time, three 10 s windows, and all 13 layers.

With 100 test clips, the 95% interval is about ±9 points. resources/baselines.csv has the chosen hyperparameters and train/valid/test accuracy. To reproduce:

uv run scripts/extract_features.py --fma-dir /path/to/fma
uv run scripts/run_baselines.py

Files

audio/<split>/<genre>/<id>.mp3  the clips, named by zero-padded FMA track id (Hub only)
audio/<split>/metadata.csv      that split's rows of metadata.csv, for the Hub viewer (Hub only)
metadata.csv                    one row per clip
checksums.sha1                  SHA-1 of every clip (sha1sum -c checksums.sha1)
LICENSE                         licenses for audio, metadata, and code
scripts/                        dataset build, duplicate detection, baselines, Hub upload
resources/                      curated exclusions, duplicate pairs, baseline results

metadata.csv columns:

column	description
track_id	FMA track id
split, genre, path	train / valid / test, class, and clip path relative to the repo root
track_title, artist_name, album_title	from FMA
track_url, artist_url, album_url	FMA pages (2017 URLs)
license, license_url, license_family	license as recorded by FMA; family is e.g. by-nc-sa
commercial_ok, derivatives_ok	the license has no NonCommercial / NoDerivatives term
tunejury_score	TuneJury score used for selection
artist_id, album_id	FMA ids (album_id is -1 if none)
fma_split, fma_subset	the track's split and subset in FMA
fma_genre_top, fma_genres	FMA's top-level genre and the track's genre tags (|-separated)
duration_full_track_s, bit_rate	duration of the full track, MP3 bit rate
sha1	SHA-1 of the clip

How it was built

scripts/build_freetzan.py selects tracks from FMA-large (deterministically):

1. Quality: keep tracks at or above the median TuneJury score.
2. License: keep Creative Commons, CC0, and Public Domain Mark.
3. Length: keep tracks with a full 30 s clip.
4. Labels: keep tracks whose genre tags all fall in one class, ignoring tags that are ancestors of another tag and a bare "Instrumental" tag.
5. Exclusions: covers, tributes, mashups, remixes, and names containing slurs (resources/exclusions.csv); "Various Artists" accounts; and duplicate recordings (resources/duplicates.csv, from scripts/find_duplicates.py, which compares log-spectrograms over time offsets), including corrupted FMA clips that share audio across unrelated tracks.
6. Splits: drawn from FMA's artist-disjoint training / validation / test splits. Artist ids that share a credited name ("X", "X ft. Y") are kept in one split.
7. Sampling: a seeded round-robin over artists in each class and split; every chosen MP3 is fully decoded to check it.

Licensing

See LICENSE. Each clip keeps its FMA license, listed in metadata.csv. Most are NonCommercial and/or NoDerivatives, and all require attribution. md[md.commercial_ok] selects the 116 clips that allow commercial use. Labels and metadata are CC BY 4.0; code is MIT.

Caveats

• Small test set. One test clip is one point of accuracy, and jazz and classical valid/test have as few as 6 distinct artists.
• Overlap with FMA. Every clip is in FMA-large (254 also in fma_small, 680 in fma_medium), so models pretrained on FMA may have heard them. FreeTZAN's splits nest inside FMA's, so FMA's training split contains no FreeTZAN valid/test clips.
• Selection bias. Only the top half of FMA by TuneJury score is eligible.
• Recording conditions. 181 clips (15%) are WFMU live radio sessions, mostly punk, rock, and folk, and much of classical is archival concert recordings, so recording conditions can correlate with labels.
• Artists and rights. Pseudonyms that share no name or recording with another artist id cannot be detected, and clips may still contain unlabeled covers or uncleared samples.

Reproducing

Requires uv, ffmpeg, FMA-large with fma_metadata/ (https://github.com/mdeff/fma), and the TuneJury scores for FMA-large (fma_large_scores_rescored.csv, not included here; available from https://github.com/yonghyunk1m/TuneJury).

uv run scripts/find_duplicates.py --fma-dir /path/to/fma --scores /path/to/fma_large_scores_rescored.csv
uv run scripts/build_freetzan.py  --fma-dir /path/to/fma --scores /path/to/fma_large_scores_rescored.csv --out .

Citation

If you use FreeTZAN, please cite FMA, from which all audio and most metadata are derived, and TuneJury, which was used to select tracks:

@inproceedings{defferrard2017fma,
  title     = {{FMA}: A Dataset for Music Analysis},
  author    = {Defferrard, Micha\"el and Benzi, Kirell and Vandergheynst, Pierre and Bresson, Xavier},
  booktitle = {Proceedings of the 18th International Society for Music Information Retrieval Conference (ISMIR)},
  year      = {2017},
  eprint    = {1612.01840},
  archivePrefix = {arXiv},
}

@article{kim2026tunejury,
  title   = {{TuneJury}: An Open Metric for Improving Music Generation Preference Alignment},
  author  = {Kim, Yonghyun and Lee, Junwon and Xia, Haiwen and Ma, Yinghao and Koo, Junghyun and Saito, Koichi and Mitsufuji, Yuki and Donahue, Chris},
  journal = {arXiv preprint arXiv:2606.17006},
  year    = {2026}
}