README.md

---
license: cc-by-4.0
task_categories:
- audio-classification
pretty_name: Beehive States (official HEAR folds)
size_categories:
- 1K<n<10K
configs:
- config_name: fold0
  data_files:
  - split: train
    path: fold0/train-*
  - split: validation
    path: fold0/validation-*
  - split: test
    path: fold0/test-*
- config_name: fold1
  data_files:
  - split: train
    path: fold1/train-*
  - split: validation
    path: fold1/validation-*
  - split: test
    path: fold1/test-*
dataset_info:
- config_name: fold0
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 48000
  - name: label
    dtype:
      class_label:
        names:
          '0': NOQUEEN
          '1': QUEEN
  splits:
  - name: train
    num_bytes: 14745623040
    num_examples: 256
  - name: validation
    num_bytes: 1843202880
    num_examples: 32
  - name: test
    num_bytes: 16588825920
    num_examples: 288
  download_size: 33178145016
  dataset_size: 33177651840
- config_name: fold1
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 48000
  - name: label
    dtype:
      class_label:
        names:
          '0': NOQUEEN
          '1': QUEEN
  splits:
  - name: train
    num_bytes: 14745623040
    num_examples: 256
  - name: validation
    num_bytes: 1843202880
    num_examples: 32
  - name: test
    num_bytes: 16588825920
    num_examples: 288
  download_size: 33178145016
  dataset_size: 33177651840
---

# Beehive States for MTEB

This artifact contains the two official full-mode Beehive States folds from
HEAR release 2021.3. Each fold contains the same 576 ten-minute, mono, 48 kHz
PCM16 NU-Hive recordings, but preserves HEAR's released preprocessing output
and cross-hive split independently. Fold 0 trains and validates on Hive 1 and
tests on Hive 3; fold 1 reverses the hives. Labels are `NOQUEEN` and `QUEEN`.

The 576 IDs and labels agree across folds, but none of the corresponding WAV
payloads are byte-identical. Consumers must therefore use the audio from the
selected configuration instead of treating the two folds as split metadata
over one shared audio table.

The source recordings are from
[NU-Hive Zenodo record 2667806](https://zenodo.org/records/2667806). The exact
preprocessed archives are from
[HEAR Zenodo record 6332517](https://zenodo.org/records/6332517):

| Fold | Archive | Size | MD5 |
| --- | --- | ---: | --- |
| fold0 | `hear2021-beehive_states_fold0-v2-full-48000.tar.gz` | 20,806,900,496 | `f9e045f9b2ddf5643edc1143304c80aa` |
| fold1 | `hear2021-beehive_states_fold1-v2-full-48000.tar.gz` | 20,806,854,397 | `8bced7ce8336bf23fd61c61f04aee109` |

`beehive_artifact_manifest.json` records the immutable source record, HEAR
preprocessing revision, split, label, and SHA-256 for every embedded WAV.

## License and citation

The source and HEAR release declare CC BY 4.0. Users must preserve attribution
to the NU-Hive and HEAR authors. Cite:

```bibtex
@inproceedings{nolasco2019beehive,
  author = {Ines Nolasco and Alessandro Terenzi and Stefania Cecchi and Simone Orcioni and Helen L. Bear and Emmanouil Benetos},
  booktitle = {2019 IEEE International Conference on Acoustics, Speech and Signal Processing},
  doi = {10.1109/ICASSP.2019.8682981},
  pages = {8256--8260},
  title = {Audio-based Identification of Beehive States},
  year = {2019}
}
```