metadata
dataset_info:
features:
- name: audio
dtype: audio
- name: id
dtype: int64
- name: labels
dtype: string
- name: sampling_rate
dtype: int64
splits:
- name: train
num_bytes: 15425963161.42
num_examples: 40966
- name: test
num_bytes: 9086607147.884
num_examples: 10231
download_size: 31287462641
dataset_size: 24512570309.304
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
HuggingFace data repo for Multi-label audio classification benchmark FSD50K. All other existing FSD50K repos on HuggingFace were either broken or didn't have the full 50k (40k train + 10k test) split
Original paper: https://arxiv.org/abs/2010.00475
Original Dataset (in .zip format): https://zenodo.org/records/4060432
from datasets import load_dataset
fsd50k = load_dataset("Chand0320/fsd50k_hf", split = "train")