load_dataset with audiofolder - specify dataset configuration via name

[maxstrobel]

Hello,

I'm struggling how to use the name parameter with load_dataset and audiofolder.

dataset = datasets.load_dataset("audiofolder", name="config_0", data_dir="./test/audio", split="train")

If I run this command, I get always the full dataset and not only the specified config - any ideas what I'm doing wrong?

---

My dataset is organized as follows:

README.md
config_0/
   - metadata.jsonl
   - 000.wav
   - 001.wav
   - ...
config_1/
   - metadata.jsonl
   - 000.wav
   - 001.wav
   - ...

with the yaml header in README.md

---
language:
- en
tags:
- audio
- size_categories:
- n<1K
source_datasets:
- original
task_categories:
  - automatic-speech-recognition
configs:
- config_name: config_0
  data_files:
  - split: train
    path: "config_0/metadata.jsonl"
- config_name: config_1
  data_files:
  - split: train
    path: "config_1/metadata.jsonl"

[bmaxdk]

Since your dataset is local and includes a README.md with a YAML configuration defining config_0 and config_1, you should load it by providing the path to the dataset folder

Can you try with this?

from datasets import load_dataset

dataset = load_dataset("./test/audio", name="config_0", split="train")

[maxstrobel]

Hi @bmaxdk,
thanks for your answer 👍

I tried that already before, but then I've the problem that the audio data is not loaded and I get a dataset with file names and the text annotations:

>>> from datasets import load_dataset

>>> dataset = load_dataset("./test/audio", name="config_0", split="train")
>>> dataset
Dataset({
    features: ['file_name', 'text'],
    num_rows: 50
})

>>> dataset[0]
{'file_name': '000.wav', 'text': 'some random text...'}

with audiofolder the audio samples are loaded and I can directly use them.
Do you know how I could add the data loading to your approach?