oliveirabruno01/openfarm-catmeows

OpenFARM CatMeows Audio Context

This prepared dataset derives from CatMeows: A Publicly-Available Dataset of Cat Vocalizations.

Source: https://zenodo.org/records/4008297 DOI: 10.5281/zenodo.4008297 Source title: CatMeows: A Publicly-Available Dataset of Cat Vocalizations Source license: CC-BY-4.0 Prepared: 2026-05-17

Scope

This is a feline audio context-classification dataset. Given a short cat vocalization, predict the eliciting context:

• brushing
• isolation_unfamiliar_environment
• waiting_for_food

These labels are source elicitation contexts. They are not clinical labels, pain scores, stress diagnoses, or validated welfare outcomes. isolation_unfamiliar_environment can support analysis of an unfamiliar-isolation recording condition, but model outputs should not be described as direct diagnoses of stress, pain, disease, or overall welfare from this dataset alone.

Source Processing Summary

• Main source WAV files: 440
• Cats: 21
• Owners: 12
• Total audio duration: 13.43 minutes
• Source label counts: {'brushing': 127, 'waiting_for_food': 92, 'isolation_unfamiliar_environment': 221}
• Audio format: mono WAV, 8 kHz, 16-bit PCM

Data Shape

Each public row is one prepared audio clip plus minimal metadata for grouping and auditing. The environment evaluating this dataset is responsible for mapping these columns into a multimodal prompt. 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 with the prepared WAV file
• audio_filename: opaque prepared filename, not the source filename
• context: target context label, one of brushing, isolation_unfamiliar_environment, or waiting_for_food
• cat_id, owner_id: source grouping IDs for cat-heldout and owner-aware leakage auditing; keep these out of primary model prompts unless running an explicit metadata ablation
• breed, breed_code, sex_status, sex_code, recording_session, vocalization_counter: parsed source metadata for analysis and ablations; these fields may correlate with source collection structure and should not be treated as neutral clinical features
• duration_sec, sample_rate_hz, channels, sample_width_bytes, frames: basic audio metadata
• rms_energy, peak_abs_amplitude, zero_crossing_rate, spectral_centroid_hz: lightweight acoustic summary features used by the baseline audit
• audio_sha256: hash of the prepared audio file
• source_url, source_doi, license: source attribution fields

Leakage Policy

Original CatMeows filenames encode the context label, cat ID, breed, sex, owner, session, and vocalization counter. Prepared media files are therefore renamed to opaque filenames such as catmeows_0001.wav.

The public rows include the target context column for scoring and parsed metadata for audits. Primary environment prompts should provide the audio and task options, not the target label, original filename metadata, or metadata fields that trivially identify or strongly proxy the context.

Recommended prompt framing: ask for the most likely source dataset context from the fixed options, and state that the task is not a clinical or welfare diagnosis.

Splits

Splits are cat-heldout at the raw train/test boundary. No cat_id appears in both raw train and raw test.

Prepared splits:

• train: class-balanced view sampled from the cat-heldout raw train set; counts {'waiting_for_food': 67, 'brushing': 67, 'isolation_unfamiliar_environment': 67}
• test: class-balanced view sampled from the cat-heldout raw test set; counts {'isolation_unfamiliar_environment': 25, 'brushing': 25, 'waiting_for_food': 25}
• train_raw: natural cat-heldout train distribution; counts {'brushing': 93, 'waiting_for_food': 67, 'isolation_unfamiliar_environment': 162}
• test_raw: natural cat-heldout test distribution; counts {'brushing': 34, 'waiting_for_food': 25, 'isolation_unfamiliar_environment': 59}

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:

• overall accuracy
• balanced accuracy
• macro F1
• per-class precision/recall/F1
• confusion matrix
• majority-class baseline
• shallow acoustic-feature baseline from baseline_audit.json

Limitations

The labels are eliciting contexts, not direct welfare or pain labels. The dataset is small and contains repeated calls from the same cats, so subject-heldout evaluation is important.