oliveirabruno01/openfarm-laying-hen-stress

OpenFARM Laying Hen Stress Vocalizations

This prepared dataset derives from the Zenodo record Vocalization Patterns in Laying Hens - An Analysis of Stress-Induced Audio Responses.

Source: https://zenodo.org/records/10433023 DOI: 10.5281/zenodo.10433023 Source title: Vocalization Patterns in Laying Hens - An Analysis of Stress-Induced Audio Responses Source license: CC-BY-4.0 Prepared: 2026-05-17

Scope

This is a laying-hen audio stress-context classification dataset. Given a bounded audio excerpt derived from a source recording, predict the source experimental context:

• poststress_control
• prestress_control
• poststress_treatment
• prestress_treatment

A derived binary label, stress_binary, marks only poststress_treatment as stress_response. This is useful for stress-proxy evaluation, but it should not be described as a clinical diagnosis.

Source Processing Summary

• Source MFCC segment rows: 1213
• Source audio files represented in MFCC table: 102
• Experimental day groups: 22
• Source stress-context counts: {'poststress_control': 16, 'prestress_control': 16, 'prestress_treatment': 35, 'poststress_treatment': 35}
• Audio files available in this prepared run: 102/102
• Source recording duration summary, seconds: {'min': 410.04514739229023, 'median': 1060.4418820861679, 'max': 3820.303673469388}
• Public excerpt duration summary, seconds: {'min': 15.0, 'median': 15.0, 'max': 15.0}

Audio Excerpt Policy

The original Zenodo recordings can be roughly 10-17 minutes long. Public HF audio in this prepared dataset is not the full source recording. Each public audio artifact is a deterministic, opaque 15s ceiling excerpt written as 16000 Hz mono WAV/PCM_16.

The default excerpt strategy is even_windows with 5 windows. Because the source MFCC table does not provide event timestamps, this samples several positions across the source recording instead of assuming the informative stress-related vocalization is at the beginning or center.

Data Shape

Each public row represents one source audio recording listed by the MFCC table, paired with one bounded public audio excerpt. DatasetDict split membership is stored by split name, so the split column is not included inside each public split.

Main fields:

• audio: HF Audio feature containing the bounded public excerpt
• audio_filename: opaque prepared filename, not the source filename
• source_audio_id: opaque stable source-recording ID
• source_group_id: opaque grouping ID used for leakage-aware analysis
• species: always laying_hen
• arm_code, experiment_group, stressor, week, day, phase, clip_index, experimental_day_group: parsed source experimental metadata for audits and ablations
• mfcc_segment_count: number of MFCC segment rows for the source recording
• stress_context: four-class target label
• stress_binary: binary stress-proxy label; only poststress_treatment is marked stress_response
• duration_sec, sample_rate_hz, channels, audio_sha256, audio_metadata_ok, audio_metadata_backend: public excerpt metadata and tracking fields
• source_duration_sec, source_sample_rate_hz, source_channels: non-identifying source recording metadata for audit only
• excerpt_duration_sec, excerpt_max_seconds, excerpt_strategy, excerpt_window_count, excerpt_sample_rate_hz, excerpt_windows_sec: public excerpt construction metadata
• source_url, source_doi, license: attribution fields

Original source filenames and source archive members are deliberately excluded from public rows because filenames encode treatment and phase. The notebook writes a local source_mapping_private.csv for audit/reproducibility, but it is not included in public HF rows.

For the primary audio benchmark, env prompts should expose audio excerpts or otherwise budgeted audio plus answer options, not parsed metadata such as phase, experiment_group, stressor, week, or day. Those fields are kept so future envs can run source-aware audits and ablations.

Splits

Splits are experimental-day-heldout at the raw train/test boundary. No experimental_day_group appears in both raw train and raw test.

Prepared splits:

• train: class-balanced view sampled from the heldout raw train set; counts {'prestress_treatment': 12, 'poststress_control': 12, 'prestress_control': 12, 'poststress_treatment': 12}
• test: class-balanced view sampled from the heldout raw test set; counts {'prestress_control': 4, 'prestress_treatment': 4, 'poststress_treatment': 4, 'poststress_control': 4}
• train_raw: natural heldout train distribution; counts {'poststress_control': 12, 'prestress_control': 12, 'prestress_treatment': 27, 'poststress_treatment': 27}
• test_raw: natural heldout test distribution; counts {'poststress_control': 4, 'prestress_control': 4, 'poststress_treatment': 8, 'prestress_treatment': 8}

Use train/test for balanced OpenFARM env development and quick model comparisons. Use train_raw/test_raw when reporting behavior on the source distribution.

Metrics

Recommended metrics: accuracy, balanced accuracy, macro F1, per-class precision/recall/F1, confusion matrix, majority-class baseline, and shallow MFCC-feature baseline from baseline_audit.json.

Limitations

The dataset is small and experimental metadata is strongly structured. Source recordings can be long and are not single isolated calls. The bounded public excerpts are deterministic benchmark artifacts, not expert-localized event windows. This dataset is best used as a compact poultry audio stress-context probe, not as a broad poultry welfare diagnosis benchmark or clinical stress diagnosis dataset. Subject IDs for individual hens are not available, so the split holds out experimental day groups rather than animals.