QBiscuits/FEA-20K

---

[AAAI 2026] Facial-R1: Aligning Reasoning and Recognition for Facial Emotion Analysis

Dataset Summary

FEA-20K is a large-scale, fine-grained dataset for Facial Emotion Analysis (FEA), containing approximately 20,000 samples. It was created using the novel Facial-R1 framework, a three-stage training process designed to align reasoning and recognition in Vision-Language Models.

The dataset is built to support explainable AI by breaking down emotion analysis into three distinct but interrelated sub-tasks. It was generated with a low-cost iterative process, starting from only 300 high-quality seed samples and using reinforcement learning to synthesize a large, high-quality corpus.

Supported Tasks

The dataset is designed to benchmark models on three core tasks:

• Facial Emotion Recognition: Classifying the primary emotion of a facial image (e.g., "disgust", "happiness").
• Facial Action Unit (AU) Recognition: Detecting the presence of specific facial muscle movements (e.g., AU4: brow lowerer).
• AU-based Emotion Reasoning: Generating natural language explanations that link the detected AUs to the final emotion prediction, explaining why a certain emotion was recognized.

Dataset Structure

• Total Samples: ~20,000
• Training Set: 17,737 samples automatically constructed via the Facial-R1 synthesis strategy.
• Test Set: 1,688 high-quality samples that have been manually verified for accuracy.

Citation

If you use this dataset in your research, please cite the original paper:

@misc{wu2025facialr1aligningreasoningrecognition,
      title={Facial-R1: Aligning Reasoning and Recognition for Facial Emotion Analysis}, 
      author={Jiulong Wu and Yucheng Shen and Lingyong Yan and Haixin Sun and Deguo Xia and Jizhou Huang and Min Cao},
      year={2025},
      eprint={2511.10254},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.10254}, 
}