README.md

---
license: other
task_categories:
  - audio-classification
pretty_name: PFANN Audio Dataset
configs:
  - config_name: default
    data_files:
      - split: audioset
        path: "data/audioset/audio/*.tar"
      - split: fma_large
        path: "data/fma_large/audio/*.tar"
      - split: micirp
        path: "data/micirp/audio/*.tar"
tags:
  - audio
  - music
  - audio-fingerprinting
  - pfann
  - webdataset
---

# Yougen/pfann_dataset

Pfann-style audio dataset, packed as **WebDataset tar shards**.

## Layout

```
data/
  audioset/
    metadata.csv
    audio/
      audioset-000.tar
      ...
  fma_large/
    metadata.csv
    audio/
      fma_large-000.tar
      ...
  micirp/
    metadata.csv
    audio/
      micirp-000.tar
      ...
```

Shard counts:

- `audioset`: 3 tar shard(s)
- `fma_large`: 136 tar shard(s)
- `micirp`: 1 tar shard(s)

Inside each tar, every sample is a pair sharing a unique key:
```
<key>.<ext>     # raw audio bytes (ext == wav / mp3 / flac / ... as in source)
<key>.json      # {"audio_id":..., "subset":..., "rel_path":..., "duration":..., "sample_rate":..., "channels":...}
```

`metadata.csv` columns:
`key, shard, audio_id, subset, rel_path, duration, sample_rate, channels`

## Loading

```python
from datasets import load_dataset

ds = load_dataset("Yougen/pfann_dataset")
print(ds)
print(ds["audioset"][0])
# sample keys: 'wav'/'mp3'/... (decoded audio), 'json' (metadata), '__key__', '__url__'
```

For streaming (no full download needed):

```python
ds = load_dataset("Yougen/pfann_dataset", streaming=True)
for example in ds["audioset"]:
    print(example["__key__"], example["json"]["audio_id"])
    break
```

HuggingFace's `webdataset` builder will automatically pair `<key>.<audio_ext>`
with `<key>.json` inside every tar and decode the audio.