README.md

---
license: cc-by-nc-4.0
language:
- en
pretty_name: Stark
tags:
- multi-modal dialogue
annotation_creators:
- machine-generated
task_ids:
- conversational
dataset_size: 93.6K
---

# Dataset Card for Stark

[🏠 Homepage](https://stark-dataset.github.io/) | [💻 Github](https://github.com/passing2961/Stark) | [📄 Arxiv](https://arxiv.org/abs/2407.03958) | [📕 PDF](https://arxiv.org/pdf/2407.03958)

## List of Provided Model Series
- **Ultron-Summarizer-Series:** [🤖 Ultron-Summarizer-1B](https://huggingface.co/passing2961/Ultron-Summarizer-1B) | [🤖 Ultron-Summarizer-3B](https://huggingface.co/passing2961/Ultron-Summarizer-3B) | [🤖 Ultron-Summarizer-8B](https://huggingface.co/passing2961/Ultron-Summarizer-8B)
- **Ultron 7B**: [🤖 Ultron-7B](https://huggingface.co/passing2961/Ultron-7B)

> 🚨 Disclaimer: All models and datasets are intended for research purposes only.

## Dataset Description
- **Repository:** [Code](https://github.com/passing2961/Stark)
- **Paper:** [Stark: Social Long-Term Multi-Modal Conversation with Persona Commonsense Knowledge](https://arxiv.org/abs/2407.03958)
- **Point of Contact:** [Young-Jun Lee](mailto:yj2961@kaist.ac.kr)

## Dataset Summary

**Stark** is a publicly available, large-scale, long-term multi-modal conversation dataset that encompasses a diverse range of social personas, multi-modality formats, time intervals, and images. To automatically construct Stark, we introduce a novel multi-modal contextualization framework, **MCU**, which generates long-term multi-modal dialogues distilled from ChatGPT and our proposed **Plan-and-Execute Image Aligner**. An overview of MCU and an example from Stark are illustrated below.

![MCU Pipeline](stark_mcu_overview.PNG)

This repository contains virtual human face images generated from text-to-image generative model (i.e., [SDXL-Lightning](https://huggingface.co/ByteDance/SDXL-Lightning)).
The virtual human face is created based on a predefined human attribute collection from [CosmicMan](https://openaccess.thecvf.com/content/CVPR2024/papers/Li_CosmicMan_A_Text-to-Image_Foundation_Model_for_Humans_CVPR_2024_paper.pdf), with the full human attribute information presented in the Appendix A.3 of our paper.

## Dataset Structure

Since the number of images is large (roughly 899K), we store and provide the image datasets in WebDataset format for efficiency.

| Field   | Type   | Description  |
|---------|--------|--------------|
| `key`   | str    | A unique identifier for each data entry in the dataset. |
| `url`   | str    | The URL path to the image stored in the dataset repository on HuggingFace. All URLs point to the base HuggingFace repository where images are stored. |
| `jpg`   | image  | The image data associated with each entry, displayed as a thumbnail in the dataset viewer. This column stores the actual image content relevant to the conversation in the dataset. |
| `json`  | dict   | Contains additional metadata and information for each image, structured as a dictionary. The JSON field typically includes the following keys: `image_source`, `image_url`, `face_description`, and `index`. |

- **`image_source`**: Source of the image (`t2i`), where:
  - `"t2i"`: Image from a general text-to-image generative model (i.e., [SDXL-Lightning](https://huggingface.co/ByteDance/SDXL-Lightning)).
- **`image_url`**: External URL where the image was originally sourced. In this dataset, the `image_url` is always an empty string.
- **`face_description`**: The virtual human face description that is created based on a predefined human attribute collection from [CosmicMan](https://openaccess.thecvf.com/content/CVPR2024/papers/Li_CosmicMan_A_Text-to-Image_Foundation_Model_for_Humans_CVPR_2024_paper.pdf).
- **`index`**: A unique index identifier for each image within the dataset, which is the same as the `key` field.

### Data Instance

The following is an example instance from the dataset.

```json
{
  "__key__": "14c9849f-ec2b-4b96-9dae-71645106c8ab",
  "__url__": "hf://datasets/passing2961/stark-face-image@c3a0b98979522e79174676583528c3c2f54741a8/stark-train-000000-of-000002.tar",
  "jpg": "<image/png>",
  "json": {
    "face_description": "A nearly full-body shot, a 85-years-old female from Japan, fit, a brown wall, ponytail blonde above chest hair, off-shoulder short scoop neckline sleeveless silk graphic dress, striped green scoop neckline normal sleeveless cotton graphic camisole, cotton graphic tie.",
    "image_source": "t2i",
    "image_url": "",
    "index": "14c9849f-ec2b-4b96-9dae-71645106c8ab"
  }
}
```

## Dataset Construction

We construct the **Stark** dataset using our proposed novel framework, MCU, which distills long-term multi-modal dialogue from ChatGPT and our proposed Plan-and-Execute Image Aligner, powered by a personalized text-to-image generative model (i.e., PhotoMaker), image database retrieval, and web search. All prompt templates used for dataset construction are presented in the Appendix of our paper.

## Languages

- English

## Further Details and Limitations

For additional information and limitations, please refer to our [paper](https://arxiv.org/abs/2407.03958).

## License and Recommendations

The **Stark** dataset is intended for research purposes only. Despite our efforts to generate high-quality and diverse personalized images, users should be mindful of ethical considerations when utilizing the dataset.

## Acknowledgements

This work was supported by a grant from the KAIST-KT joint research project through AI Tech Lab, Institute of Convergence Technology, funded by KT [Project No. G01230605, Development of Task-oriented Persona-based Dialogue Generation Combining Multi-modal Interaction and Knowledge Modeling].

## Citation

If you find the resources in this repository useful, please cite our work:

```
@article{lee2024stark,
  title={Stark: Social Long-Term Multi-Modal Conversation with Persona Commonsense Knowledge},
  author={Lee, Young-Jun and Lee, Dokyong and Youn, Junyoung and Oh, Kyeongjin and Ko, Byungsoo and Hyeon, Jonghwan and Choi, Ho-Jin},
  journal={arXiv preprint arXiv:2407.03958},
  year={2024}
}
```