oliveirabruno01/openfarm-zoo-valence-arousal

OpenFARM Zoo Valence Arousal

This prepared dataset derives from the Figshare collection Emotion recognition study, associated with Hiisivuori et al. (2025), Human recognition of emotional valence and arousal of zoo animals.

Source: https://doi.org/10.6084/m9.figshare.c.7807931 Article: https://www.nature.com/articles/s41598-025-28646-7 Article DOI: 10.1038/s41598-025-28646-7 Source license: CC BY 4.0 according to the Figshare articles Prepared: 2026-05-18

Scope

This is a tiny eval-only visual affect benchmark. Given a short zoo-animal visual stimulus, predict expert-labeled emotional valence and/or arousal.

• —Species counts: {'Siberian tiger': 5, 'Turkmenian markhor': 5, 'Barbary macaque': 5}
• —Rows: 15 eval examples
• —Split: test only
• —Valence labels: {'negative': 6, 'neutral': 3, 'positive': 6}
• —Arousal labels: {'high': 9, 'low': 6}

These are expert emotional-state labels from the source study. They are not clinical pain scores or a complete welfare diagnosis.

Prepared Media

The Figshare source provides three video compilations. Each row here is clipped from the compilation using black-frame separator detection and preserves the full stimulus segment used in the human study: real-speed clip, slow-motion clip, and still-frame portion. The primary video column is muted because the paper describes the study as visual recognition from mute clips.

Each row also includes a compact 3x3 filmstrip image for image-model evaluation.

Audio audit summary: {'activeaudiosecmin': 8.02, 'activeaudiosecmedian': 15.42, 'activeaudiosecmax': 22.13, 'activeaudioratiomedian': 0.481}

Audio-bearing video exported in this dataset build: False. If present, video_with_audio should be treated as an explicit multimodal ablation, not the headline human-comparable benchmark.

Data Shape

Main fields:

• —filmstrip: HF Image feature with a 3x3 visual summary
• —video: HF Video feature with the muted prepared MP4 stimulus
• —video_with_audio: optional HF Video feature, present only when EXPORT_AUDIO_VARIANT=True
• —species: target animal species
• —valence: expert label, one of negative, neutral, positive
• —arousal: expert label, one of low, high
• —valence_arousal: combined helper label
• —duration_sec: prepared stimulus duration
• —source_url, source_article_url, source_doi, article_doi, license: source attribution fields

Leakage Notes

Public rows use opaque media filenames and intentionally omit clip codes, source filenames, segment timestamps, expert interpretation notes, pre-formatted messages, task ids, and OpenFARM-specific row ids. The environment should build prompts from the media and selected non-target fields only.

Sources

• —Figshare collection: https://doi.org/10.6084/m9.figshare.c.7807931
• —Article: https://www.nature.com/articles/s41598-025-28646-7
• —Article DOI: 10.1038/s41598-025-28646-7