README.md

---
license: cc-by-nc-4.0
task_categories:
- other
tags:
- multimodal
- personality-understanding
- mbti
- fairness
---

# DMSP — Dataset for Multimodal Personality Research

The **DMSP** (Demographic-annotated Multimodal Student Personality) dataset is a resource designed to address the challenges in personality detection from multimodal content, particularly focusing on the **Myers-Briggs Type Indicator (MBTI)** and **fairness evaluation**. 

Unlike existing datasets that rely heavily on text-only inputs, DMSP integrates **Visual, Audio, and Textual modalities**. By incorporating fairness attributes (Gender, Age, Race) and continuous soft labels, this dataset offers a more accurate reflection of how personality traits manifest in real-world scenarios.

- **Paper**: [Debiased Multimodal Personality Understanding through Dual Causal Intervention](https://huggingface.co/papers/2605.06371)
- **GitHub Repository**: [Sabrina-han/DCAN](https://github.com/Sabrina-han/DCAN)

## Key Features
* **Multimodal Integration**: Leverages CLIP (Visual), Wav2Clip (Audio), and CLIP Sentence Embeddings (Text) for robust feature representation.
* **Fairness-Oriented**: Includes demographic annotations (Gender, Age, Race) to facilitate fairness analysis and bias mitigation in AI models.
* **Soft Labeling**: Utilizes continuous scores for the 4 MBTI dimensions (E/I, N/S, F/T, J/P), moving beyond the limitations of hard binary classifications.

## Dataset Structure

```text
DMSP/
├── train.csv                 # Training labels and metadata
├── test.csv                  # Test labels and metadata
├── train_clipimage.pkl       # Visual features for training set (CLIP ViT-B/32)
├── test_clipimage.pkl        # Visual features for test set
├── train_audio_wav2clip.pkl  # Audio features for training set (Wav2Clip)
├── test_audio_wav2clip.pkl   # Audio features for test set
├── train_clipsentence.pkl    # Text features for training set (CLIP Sentence)
└── test_clipsentence.pkl     # Text features for test set
```

## Sample Usage

You can load the dataset using the following snippet found in the official repository:

```python
from train_FMPD_MBTI_baseline_fixed import DMSPDataset

train_ds = DMSPDataset(
    csv_file='DMSP/train.csv',
    data_dir='DMSP',
    split='train'
)

sample = train_ds[0]
print(sample.keys())
# Output: ['vid', 'mbti', 'demo', 'visual', 'audio', 'text']
```

## Citation

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

```bibtex
@article{han2024debiased,
  title={Debiased Multimodal Personality Understanding through Dual Causal Intervention},
  author={Li, Han and others},
  journal={arXiv preprint arXiv:2605.06371},
  year={2024}
}
```