dolphinteam/OpenWhistle-Classification-Finetuning

OpenWhistle Classification Finetuning Dataset

dolphinteam/OpenWhistle-Classification-Finetuning is the public classification finetuning dataset used for dolphin whistle identity classification. It contains short whistle clips, whistle-level metadata, fundamental-frequency tracks, rendered F0 spectrograms, and integer class labels.

The dataset has two configs:

• balanced: the main six-class benchmark, with exactly 500 clips per class.
• unbalanced: all 8,354 expert-annotated whistles across ten classes, with the natural class imbalance.

Both configs are split by recording session, so no session appears in more than one of train, validation, or test.

Dataset Contents

• Hugging Face repo: dolphinteam/OpenWhistle-Classification-Finetuning
• Configs: balanced (main benchmark) and unbalanced
• Audio: mono, 44.1 kHz
• Public columns common to both configs: audio, label, name, onset, offset, duration, recording_duration, whistle_type, whistle_name, f0_time, f0_hz, f0_conf, f0_ok, f0_bad_reason, f0_spectrogram
• The unbalanced config additionally includes snr_db, the estimated clip-level signal-to-noise ratio in dB.

Balanced Config

The balanced config contains six classes:

• NSW_1 (label=0)
• SW_Luna (label=1)
• SW_Nana (label=2)
• SW_Neo (label=3)
• SW_Nikita (label=4)
• SW_Yosefa (label=5)

Split	Rows	NSW_1	SW_Luna	SW_Nana	SW_Neo	SW_Nikita	SW_Yosefa	Sessions
train	2,100	350	350	350	350	350	350	161
validation	450	75	75	75	75	75	75	54
test	450	75	75	75	75	75	75	37
Total	3,000	500	500	500	500	500	500	252

The split assignment was generated with seed 42 and exact class balancing. Session leakage checks found no overlap between any pair of splits.

Unbalanced Config

The unbalanced config contains ten classes:

• NSW_3 (label=0)
• NSW_2 (label=1)
• NSW_1 (label=2)
• SW_Dana (label=3)
• SW_Luna (label=4)
• SW_Nana (label=5)
• SW_Neo (label=6)
• SW_Nikita (label=7)
• SW_Shy (label=8)
• SW_Yosefa (label=9)

Note that the integer labels differ from those of the balanced config.

Split	Rows	NSW_3	NSW_2	NSW_1	SW_Dana	SW_Luna	SW_Nana	SW_Neo	SW_Nikita	SW_Shy	SW_Yosefa	Sessions
train	5,848	31	19	460	235	2,054	365	1,567	622	57	438	145
validation	1,253	7	4	99	50	440	78	336	133	12	94	54
test	1,253	7	4	99	50	440	78	336	133	12	94	62
Total	8,354	45	27	658	335	2,934	521	2,239	888	81	626	261

No recording session appears in more than one split. The rare classes (NSW_3, NSW_2, SW_Shy) have few test examples, so per-class results for them should be interpreted with care.

Loading The Data

from datasets import load_dataset

balanced = load_dataset(
    "dolphinteam/OpenWhistle-Classification-Finetuning",
    "balanced",
)
unbalanced = load_dataset(
    "dolphinteam/OpenWhistle-Classification-Finetuning",
    "unbalanced",
)

Earlier Revisions

In earlier revisions of this repository, the current unbalanced config was named all; its data files are unchanged, so results reported for all correspond to unbalanced. Those revisions also contained a smaller ten-class config with capped class sizes, then named unbalanced (3,488 clips), and *-review-sample configs for inspection. They remain available by revision:

from datasets import load_dataset

previous = load_dataset(
    "dolphinteam/OpenWhistle-Classification-Finetuning",
    "all",
    revision="364e3d31b5a96b0a524d0a5d18a46902f3f77020",
)

License

This dataset is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0), as recorded in its Croissant/RAI metadata.

Paper, collection, and authors

• Paper: OpenWhistle: A Large-Scale Longitudinal Dataset and Benchmark of Bottlenose Dolphin Vocalizations
• Hugging Face collection: NeurIPS 2026 OpenWhistle
• This repository: https://huggingface.co/datasets/dolphinteam/OpenWhistle-Classification-Finetuning

Authors: Faadil Mustun, Chiara Semenzin, Roberto Dessi, Pablo Robin Guerrero, Pierre Orhan, Alexis Emanuelli, Emanuele Rossi, Yair Lakretz, Gonzalo de Polavieja, and German Sumbre.

Citation

@misc{mustun2026openwhistle,
  title         = {OpenWhistle: A Large-Scale Longitudinal Dataset and Benchmark of Bottlenose Dolphin Vocalizations},
  author        = {Mustun, Faadil and Semenzin, Chiara and Dessi, Roberto and Robin Guerrero, Pablo and Orhan, Pierre and Emanuelli, Alexis and Rossi, Emanuele and Lakretz, Yair and de Polavieja, Gonzalo and Sumbre, German},
  year          = {2026},
  eprint        = {2609.34839},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2609.34839}
}