EmoWorld-130K

Dataset Details

Description

EmoWorld-130K is a large-scale human-centric multimodal affective reasoning dataset designed to advance the development of emotion-aware multimodal models or world models.

Unlike existing emotion datasets that primarily focus on perception-level recognition, EmoWorld-130K provides a unified benchmark covering the complete affective intelligence pipeline, from emotion perception, to affective understanding, and finally to empathetic interaction.

The dataset contains approximately 130K multimodal instruction-following samples spanning eight core affective tasks:

• Multimodal Sentiment Analysis (MSA)
• Basic Multimodal Emotion Recognition (B-MER)
• Open-Vocabulary Multimodal Emotion Recognition (OV-MER)
• Multimodal Intent Recognition (MIR)
• Multimodal Humor Understanding (MHU)
• Multimodal Sarcasm Understanding (MSU)
• Empathetic Response Generation (ERG)
• Emotional Support Conversation (ESC)

Each sample is equipped with structured and explicit reasoning trajectories, enabling models to learn not only the final prediction or response generation, but also the underlying affective reasoning process.

EmoWorld-130K is constructed through a three-stage pipeline:

1. Reasoning trajectory distillation from expert multimodal language models to obtain high-quality affective reasoning traces.
2. Automated answer verification through reverse reasoning-based consistency checking.
3. Rigorous manual quality control to ensure annotation reliability and reasoning quality.

By integrating perception, understanding, and interaction-level tasks into a single unified dataset, EmoWorld-130K aims to facilitate the training and evaluation of next-generation emotion-aware multimodal AI systems.

• Curated by: [To be continued]

Dataset Sources

• Repository: https://github.com/waHAHJIAHAO/OneEmo
• Paper: OneEmo: A Unified Multimodal Reasoning Model for Emotion Perception, Understanding, and Interaction

Uses

Direct Use

EmoWorld-130K is intended for research purposes in multimodal affective intelligence, including:

• Training emotion-aware multimodal large language models.
• Developing models capable of affective perception, reasoning, and empathetic interaction.
• Evaluating multimodal emotion understanding across diverse affective tasks.
• Studying chain-of-thought-style affective reasoning and explainable multimodal AI.

The dataset can be used for supervised fine-tuning (SFT), reinforcement learning from human/AI feedback (RLHF/RLAIF), instruction tuning, and benchmark evaluation of multimodal foundation models.

Dataset Structure

Each instance in EmoWorld-130K contains multimodal inputs and structured affective reasoning supervision.

A typical sample includes:

• Multimodal input: image/video content and associated conversational context.
• Instruction: task-specific user query.
• Reasoning trajectory: explicit intermediate affective reasoning process.
• Response/Answer: final prediction, explanation, or generated empathetic response.
• Task category: one of the eight affective tasks.

The dataset is organized according to three affective intelligence levels:

Level	Description	Tasks
Perception-level	Recognizing observable affective signals	MSA, B-MER, OV-MER
Understanding-level	Reasoning about intentions, humor, sarcasm, and emotional states	MIR, MHU, MSU
Interaction-level	Generating socially appropriate emotional responses	ERG, ESC

Each sample is designed to bridge the gap between affective perception and human-like emotional interaction.

Dataset Creation

Curation Rationale

Existing multimodal emotion datasets mainly focus on isolated recognition tasks and lack explicit reasoning supervision. Although recent multimodal large language models demonstrate strong visual-language understanding, their ability to perceive, interpret, and respond to human emotions remains limited.

EmoWorld-130K is created to address this limitation by providing large-scale, explainable, and unified affective reasoning data that supports the development of emotionally intelligent multimodal models.

Who are the source data producers?

The original multimodal data sources are collected from publicly available datasets and human-centric multimedia resources.

The original creators and contributors of these datasets retain their respective ownership and licenses.

Annotations

Annotation process

Annotations in EmoWorld-130K include:

• Task-specific affective labels.
• Structured reasoning trajectories.
• Final answers or generated responses.

The annotation pipeline combines:

• Expert model-based reasoning generation.
• Automated consistency verification.
• Manual quality inspection.

This hybrid approach improves scalability while maintaining annotation quality.

Who are the annotators?

Annotations are produced through a combination of:

• Expert multimodal foundation models.
• Human reviewers responsible for quality verification.

Additional annotator information will be provided when available.

Recommendations

Users should evaluate model behavior carefully before deploying emotion-aware systems.

Recommended practices include:

• Performing domain-specific evaluation.
• Considering cultural and demographic diversity.
• Avoiding high-impact decisions based solely on model-generated emotional interpretations.
• Combining model outputs with human judgment in sensitive scenarios.

Citation

If you use EmoWorld-130K in your research, please cite:

BibTeX:

@article{huang2026oneemo,
      title={OneEmo: A Unified Multimodal Reasoning Model for Emotion Perception, Understanding, and Interaction}, 
      author={Jiahao Huang and Zheng Lian and Jingyi Zhang and Zhide Chen and Xiaojiang Peng and Shaonan Wang},
      year={2026},
      journal={arXiv preprint arXiv:2608.06013},
}