OpenWhistleNeurIPS26/OpenWhistle-Pretraining

OpenWhistle Pretraining Dataset

OpenWhistleNeurIPS26/OpenWhistle-Pretraining is the public unlabeled audio dataset used for OpenWhistle pretraining. It contains 96 kHz dolphin acoustic segments with timing and recording metadata, but no whistle/noise labels.

The main default config is the complete pretraining dataset. A smaller deterministic review-sample config is also provided so reviewers can inspect representative examples quickly without downloading the full dataset.

Dataset contents

• Hugging Face repo: OpenWhistleNeurIPS26/OpenWhistle-Pretraining
• Public columns: audio, start_time, end_time, duration, year, hydrophone
• Sampling rate: 96 kHz, mono audio
• Labels: none; this dataset is intended for unsupervised or self-supervised pretraining

Full dataset splits

Split	Rows	Duration (s)	Duration (h)
train	28,410	367,792.80	102.165
validation	3,370	43,629.20	12.119
Total	31,780	411,422.00	114.284

Full dataset coverage by year

Split	2019	2020	2021	2023	2024
train	920	1,420	14,376	9,885	1,809
validation	119	179	1,820	1,055	197
Total	1,039	1,599	16,196	10,940	2,006

Full dataset coverage by hydrophone

Split	channel_0	channel_1	channel_2
train	20,808	7,130	472
validation	2,599	741	30
Total	23,407	7,871	502

Full dataset coverage by year and hydrophone

Split	2019 ch0	2020 ch0	2021 ch0	2023 ch0	2023 ch1	2023 ch2	2024 ch0	2024 ch1
train	920	1,420	14,376	3,982	5,431	472	110	1,699
validation	119	179	1,820	464	561	30	17	180
Total	1,039	1,599	16,196	4,446	5,992	502	127	1,879

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 pretraining dataset while keeping the download small enough for quick manual inspection. Because the pretraining dataset is unlabeled, the sample is not class-balanced. Instead, it preserves the original train and validation split names and uses the same audio and metadata columns as the full dataset.

The target size was set to 480 rows, following the same review-sample size used for the OpenWhistle CNN dataset. Rows were allocated across splits to keep the same approximate train/validation ratio as the full pretraining dataset: 430 examples for train and 50 examples for validation.

For each split, the dataset was streamed from OpenWhistleNeurIPS26/OpenWhistle-Pretraining, the audio column was cast to Audio(sampling_rate=96000, mono=True, decode=False), and rows were shuffled with a fixed seed and a finite shuffle buffer before taking the requested number of examples. Sampling used base seed 42, shuffle buffer size 64, and split-specific seeds derived from the split order (train=42, validation=43). The resulting config is named review-sample.

Review sample size

Split	Rows	Duration (s)	Duration (h)
train	430	6,017.60	1.672
validation	50	642.40	0.178
Total	480	6,660.00	1.850

Review sample coverage

Split	Year counts	Hydrophone counts
train	2019: 12; 2020: 33; 2021: 218; 2023: 137; 2024: 30	channel_0: 336; channel_1: 93; channel_2: 1
validation	2021: 20; 2023: 30	channel_0: 37; channel_1: 11; channel_2: 2

Loading the data

from datasets import load_dataset

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