cirimus/super-emotion

Super Emotion Dataset

Dataset Summary

The Super Emotion Dataset is a large-scale dataset for emotion classification, aggregated from multiple sources:

• MELD
• GoEmotions
• Carer
• ISEAR
• SemEval
• Crowdflower

It contains 519,812 total samples, respecting original train/validation/test splits where possible. It supports 7 emotion categories which had maximum support in the aggregation: joy, sadness, anger, fear, love, neutral, surprise. Note that we merged some categories to this end (happiness and joy, hate and anger, grief and sadness).

Supported Tasks

This dataset is designed for emotion classification and can be used for:

• Single-label classification
• Multi-label emotion recognition
• Fine-tuning language models

Dataset Structure

The dataset follows the structure:

Column	Type	Description
text	string	The input text
label	string	The assigned emotion label
source	string	The original dataset

Splits:

• Train: 412,059 samples
• Validation: 51,443 samples
• Test: 56,310 samples

Class distribution by source:

Source	NEUTRAL	SURPRISE	FEAR	SADNESS	JOY	ANGER	LOVE	All
MELD	4,710	1,205	268	683	1,743	1,109	0	9,718
Emotion Dataset	0	572	1,937	4,666	5,362	2,159	1,304	16,000
ISEAR	0	11,978	38,170	96,949	112,853	45,854	27,643	333,447
GoEmotions	14,219	1,060	596	1,403	1,452	1,567	2,086	22,383
Crowdflower	6,910	1,750	0	4,132	4,167	1,146	3,074	21,179
SemEval	0	361	1,242	2,008	2,477	2,544	700	9,332
All	25,839	16,926	42,213	109,841	128,054	54,379	34,807	412,059

Citation

If you use this dataset, please cite the original sources (Crowdflower 2016, Elvis et al. 2018, Demszky et al. 2020, Vikash 2018, Poria et al. 2019, EI-reg Mohammad et al. 2018) as well as:

@inproceedings{JdFE2025d,
  title = {The Super Emotion Dataset},
  author = {Enric Junqu\'e de Fortuny},
  year = {2025},
  howpublished = {\url{https://huggingface.co/cirimus/super-emotion}},
}