CREMA-D-Irony

A controllable irony extension of CREMA-D, introduced in SynIB: An Informational Bottleneck for Maximizing Synergy in Multimodal Learning (arXiv:2606.09853 · code: github.com/kkontras/SynIB).

For a fraction α of clips, the original video is kept but its audio is replaced with a donor clip carrying a contradicting emotion, and the sample is relabeled SAR (ironic). The label then becomes recoverable only by combining audio and video — a controllable amount of cross-modal synergy.

Configs

config	what it is	media	folds
alpha_0	plain CREMA-D (no irony) — our 5-fold CV splits, labels = original emotion	splits only (bring CREMA-D)	0–4 (all 5)
alpha_0.1 / alpha_0.5 / alpha_1.0 (default) / alpha_2.0	the irony rates used in the paper; self-contained (decoded audio + video frames embedded)	included	0 (paper fold, seed 109)

Higher α ⇒ more ironic examples (≈ α/6 of the data). For the irony configs, the donor swap is fully deterministic from (α, seed). alpha_0 involves no irony and therefore no seed — just the fixed splits, shared for all five folds (the paper used three of them).

column	meaning
audio	waveform (22.05 kHz) — donor's audio when ironic, else the clip's own (α>0 configs)
frames	1-FPS video frames, always the original clip (α>0 configs)
label	emotion (NEU/HAP/SAD/FEA/DIS/ANG) or SAR when ironic
original_emotion	the clip's original CREMA-D emotion
is_ironic	1 if made ironic
clip / donor_clip	CREMA-D clip id, and the donor whose audio was swapped in
fold	CV fold index

from datasets import load_dataset
ds = load_dataset("kkontras/crema-d-irony", "alpha_0.5")   # self-contained: ds["test"][0]["audio"], ["frames"]
splits = load_dataset("kkontras/crema-d-irony", "alpha_0") # 5-fold CV splits over plain CREMA-D

License & attribution

Adapted from CREMA-D and released, like it, under the Open Database License (ODbL) v1.0; attribute CREMA-D and keep its license notice.

Citation

@article{kontras2026synib, title={SynIB: An Informational Bottleneck for Maximizing Synergy in Multimodal Learning},
  author={Kontras, Konstantinos and Popordanoska, Teodora and Strypsteen, Thomas and Chatzichristos, Christos and Blaschko, Matthew and De Vos, Maarten and Liang, Paul Pu}, journal={arXiv:2606.09853}, year={2026}}
@article{cao2014crema, title={CREMA-D: Crowd-sourced emotional multimodal actors dataset},
  author={Cao, Houwei and Cooper, David G and Keutmann, Michael K and Gur, Ruben C and Nenkova, Ani and Verma, Ragini}, journal={IEEE Transactions on Affective Computing}, year={2014}}