README.md

---
license: mit
size_categories:
- 1K<n<10K
pretty_name: EDM CUE
dataset_info:
  features:
  - name: id
    dtype: int64
  - name: title
    dtype: string
  - name: artists
    dtype: string
  - name: duration
    dtype: int64
  - name: genre
    sequence: string
  - name: key
    sequence: string
  - name: beat_grid
    struct:
    - name: bpm
      dtype: float64
    - name: init_beat
      dtype: int64
    - name: start_pos
      dtype: float64
    - name: time_sig
      dtype: string
  - name: cue_pts
    sequence: float64
  splits:
  - name: train
    num_bytes: 690103
    num_examples: 4609
  - name: valid
    num_bytes: 16842
    num_examples: 101
  download_size: 292711
  dataset_size: 706945
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: valid
    path: data/valid-*
tags:
- music
---

```
from datasets import load_dataset
captions = load_dataset("disco-eth/edm-cue")
```

# What is EDM-CUE?

The EDM-CUE dataset contains metadata for ~5k EDM tracks. Cue points are essential for DJs, so we asked the question "can they be placed by a learned system?" To Answer this question we gathered 21k cue points manually placed by human experts, and provide them in this dataset for future use.

To cite this dataset or for more information, please see [*Cue Point Estimation using Object Detection*](https://www.arxiv.org/abs/2407.06823).


## Dataset Structure

The dataset contains the following features:

* id: Track ID on Deezer.
* title: Track name.
* artists: Track artists.
* duration: Track duration in seconds.
* genre: List of genres (possibly incomplete).
* key: Alphanumeric key of the track generated by rekordbox (possibly incorrect).
* beat_grid: Contains 4 values to compute the grid. 'start_pos' is the starting position in seconds, 'init_beat' is the beat count of 'start_pos', 'bpm' is the beats per minute tempo, and 'time_sig' is the time signature in string format.
* cue_pts: Position of the cue point in seconds.