README.md

---
license: other
task_categories:
  - audio-classification
pretty_name: DaCoS Dataset
configs:
  - config_name: default
    data_files:
      - split: train
        path: "data/train/audio/*.tar"
tags:
  - audio
  - music
  - cover-song-identification
  - dacos
  - webdataset
---

# Yougen/datacos_dataset

DaCoS (Da Cover Song) audio dataset, packed as **WebDataset tar shards**.
Each row in the original `output.csv` is one recording; rows sharing the
same `clique` are alternative versions ("covers") of the same underlying
song.

## Layout

```
data/
  train/
    metadata.csv
    audio/
      train-000.tar
      train-001.tar
      ...
```

Shard counts:

- `train`: 90 tar shard(s)

Inside each tar, every sample is a pair sharing a unique key:
```
<key>.mp3       # raw MP3 bytes
<key>.json      # {"id":..., "rel_path":..., "wav_format":"mp3",
                #   "duration":..., "label_str":"<clique>", "label":<int>,
                #   "clique":..., "version":..., "title":...,
                #   "performer":..., "video_id":...}
```

`metadata.csv` columns:
`key, shard, id, rel_path, wav_format, duration, label_str, label, clique, version, title, performer, video_id`


## Loading

```python
from datasets import load_dataset

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

For streaming (no full download needed):

```python
ds = load_dataset("Yougen/datacos_dataset", streaming=True)
for example in ds["train"]:
    print(example["__key__"], example["json"]["clique"], example["json"]["title"])
    break
```

HuggingFace's `webdataset` builder will automatically pair `<key>.mp3` with
`<key>.json` inside every tar and decode the audio.