README.md

---
task_categories:
- audio-classification
tags:
- dolphin
- bioacoustics
- whistle-detection
- audio
- spectrogram
dataset_info:
  config_name: review-sample
  features:
  - name: audio
    dtype:
      audio:
        decode: false
  - name: spectrogram
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': noise
          '1': whistle
  - name: file_name
    dtype: string
  - name: recording
    dtype: string
  - name: onset
    dtype: float64
  - name: offset
    dtype: float64
  splits:
  - name: train
    num_bytes: 37497957.0
    num_examples: 376
  - name: test
    num_bytes: 10431395.0
    num_examples: 104
  download_size: 38245514
  dataset_size: 47929352.0
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
- config_name: review-sample
  data_files:
  - split: train
    path: review-sample/train-*
  - split: test
    path: review-sample/test-*
---

# OpenWhistle CNN Dataset

`OpenWhistleNeurIPS26/OpenWhistle-CNN` is the public CNN dataset used for binary
dolphin whistle detection. It contains audio windows, spectrogram images, and
binary labels:

- `noise` (`label=0`)
- `whistle` (`label=1`)

The main dataset is the complete session-disjoint dataset used for training and
evaluation. A smaller deterministic `review-sample` config is also provided so
reviewers can inspect representative examples quickly.

## Dataset contents

- Hugging Face repo: `OpenWhistleNeurIPS26/OpenWhistle-CNN`
- Public columns: `audio`, `spectrogram`, `label`, `file_name`, `recording`, `onset`, `offset`

## Full dataset splits

| Split | Rows | Noise | Whistle | Sessions | Window hours |
| --- | ---: | ---: | ---: | ---: | ---: |
| `train` | 53,828 | 26,914 | 26,914 | 195 | 5.980885 |
| `validation` | 5,980 | 2,990 | 2,990 | 26 | 0.664445 |
| `test` | 16,708 | 8,354 | 8,354 | 261 | 1.856444 |
| **Total** | **76,516** | **38,258** | **38,258** | **482** | **8.501775** |

The `train` and `validation` splits come from the non-2019/2020 pool. The
`test` split is a manual 2019-2020 test split built from the full classification
`all` config.

## Review sample

The `review-sample` config is a small deterministic subset of the same public
dataset. It was created only to make review and manual inspection easier. It is
not a replacement for the full dataset used for model development or reporting.

### How the review sample was created

The review sample was designed to preserve the structure of the full dataset
while keeping the download small enough for quick manual inspection. The sample
keeps the same binary label definition as the full dataset and preserves the
train/test separation: reviewer training examples are drawn from the original
training and validation data, while reviewer test examples are drawn only from
the original test data.

Within each reviewer split, examples were sampled separately for `noise`
(`label=0`) and `whistle` (`label=1`) so that both classes are equally
represented. This avoids a reviewer sample dominated by one class and makes it
easier to inspect positives and negatives side by side. The target sizes were
chosen to keep the same approximate train/test ratio as the full CNN dataset:
376 examples for `train` and 104 examples for `test`, for 480 examples total.

Sampling was deterministic, using seed `42`, so the same review sample can be
rebuilt exactly from the prepared public dataset. The resulting config is named
`review-sample`.

### Review sample size

| Split | Rows | Noise | Whistle | Source splits | Source rows |
| --- | ---: | ---: | ---: | --- | ---: |
| `train` | 376 | 188 | 188 | `train`, `validation` | 59,808 |
| `test` | 104 | 52 | 52 | `test` | 16,708 |
| **Total** | **480** | **240** | **240** | | |

## Loading the data

```python
from datasets import load_dataset

full = load_dataset("OpenWhistleNeurIPS26/OpenWhistle-CNN")
review = load_dataset("OpenWhistleNeurIPS26/OpenWhistle-CNN", "review-sample")
```